# verifiable.pro > verifiable.pro reports in English and Simplified Chinese on verifiable AI: evidence that lets people check what AI services claim, such as which model answered a request or which software handled it, before trusting them. Coverage starts from trusted execution environments (TEEs), remote attestation and confidential computing. It compares them with other ways to check a claim: zkML, verifiable inference and formal verification check computations and code, while AI evaluations, audits, model signing and supply-chain provenance check models and their origins. The comparison also extends to content credentials (C2PA) and watermarking, agent accountability and the verification of AI governance commitments. The site publishes news briefs, analysis, explainers and guides, a resource library and a glossary. Besides the articles, the site keeps a hub on trusted execution environments (TEEs) and a map of the verification stack: seven layers from Hardware & compute (T1) to Policy (T7), with five trust tiers. Every piece says what a check establishes and what it still leaves to trust, and the site never calls a service verified. Each factual claim in an article traces to a numbered source listed with it; in the Markdown, [^n] cites source n. Every news brief ends with a "What we checked" box that names the primary sources opened, the day they were opened and what could not be confirmed, and the site's interpretation sits apart in a "Why it matters" box. A vendor's security claim is reported as the vendor's claim. Corrections are listed with their dates at the end of the article they correct. Each article's byline gives the name of the person who wrote it or, for an article that software drafted with AI models, one of the site's automated bylines (verifiable.pro Wire, verifiable.pro Research or verifiable.pro Research), marked as automated. AI tools also translate, and an editor decides what is published and checks each claim against its source before publication. Every article and page has a Markdown source beside it: `/articles//en.md` and `zh.md` for the article at `/articles//` and `/zh/articles//`, and `/pages/.en.md` or `.zh.md` for the page at `//` and `/zh//`. The site manifest lists the SHA-256 digest of every source and data file; its `digest` field is the SHA-256 of the manifest without its `digest` and `digest_note` fields, serialized with sorted keys, no whitespace and UTF-8. To check a page, download its source, compute its SHA-256 and compare it with the source's entry in the manifest; `tools/verify_manifest.py` in the source repository checks every entry. A match shows that the text is what the site recorded when it was built, not that the reporting is correct. To cite a page, give its title, its address (not the address of its Markdown source), verifiable.pro and its filing or update date. An article with a lead image also has `/articles//illustration.json` beside its sources, with the picture's alt text and caption in both languages. For a photo, the file gives its credit, license and source, and how `tools/photos.py` in the source repository cropped it and adjusted its colors; the caption says what the photo shows, and says so when it does not show the story's own subject. For an illustration, the file holds the scene description that `tools/illustrate.py` draws the picture from; the illustrations are conceptual and depict no real event, person or product. The manifest lists that file and the pictures under `/assets/media/` with their digests. ## Start here - [Powerful AI should come with evidence](https://verifiable.pro/pages/vision.en.md): Our vision for AI whose claims people can check. - [About verifiable.pro](https://verifiable.pro/pages/about.en.md): What we cover, how we check it, and how you can check our work. - [Contributors](https://verifiable.pro/pages/contributors.en.md): People contributing to verifiable.pro. - [Privacy and site statistics](https://verifiable.pro/pages/privacy.en.md): Your choice about visit measurement, and what stays in your browser. - [What makes an AI system verifiable?](https://verifiable.pro/articles/what-makes-an-ai-system-verifiable/en.md): Step 1 of the TEE path (explainer). Verifying an AI system takes a precise claim, evidence that someone other than the developer can check, and a party able to check it. This explainer shows the layers where claims arise, then describes six kinds of evidence and the trust each one leaves in place. - [Remote attestation for AI workloads, explained](https://verifiable.pro/articles/remote-attestation-for-ai-workloads-explained/en.md): Step 2 of the TEE path (explainer). Before a client sends a prompt or a model owner releases its weights, remote attestation can show which software a machine loaded and on what hardware. This explainer shows how that evidence is produced and checked, including on GPUs, and what a valid report leaves open. - [How to read a remote attestation report](https://verifiable.pro/articles/how-to-read-a-remote-attestation-report/en.md): Step 3 of the TEE path (guide). A nine-step worksheet for deciding whether a service's hardware-signed evidence shows software and settings you are prepared to trust, before any data leaves your systems. - [Private AI Compute Services: What the Published Designs Let Outsiders Check](https://verifiable.pro/articles/private-ai-compute-published-designs/en.md): Step 4 of the TEE path (analysis). Apple, Meta, Google and Cohere now route AI requests to servers whose chips sign a record of the software they loaded. Their published designs differ in what anyone outside each company can check about that software. - [Three Ways to Check an AI Answer: Proofs, Enclaves and Re-execution Compared](https://verifiable.pro/articles/proofs-enclaves-and-re-execution-compared/en.md): Step 5 of the TEE path (analysis). Zero-knowledge proofs, attested enclaves and re-execution each give evidence about which model produced an answer. None is both cheap and low in trust assumptions, so the choice should depend on what each one establishes and what it leaves to trust. - [FLASH: checkpointing confidential LLM inference on NVIDIA H100](https://verifiable.pro/articles/flash-confidential-gpu-checkpoint/en.md): Explainer. FLASH reuses GPU ciphertext and UVM dirty-page tracking to shorten checkpoint pauses. Its H100 and Intel TDX evaluation distinguishes pause latency, serving impact, and recovery time. - [How to verify a signed model before you load it](https://verifiable.pro/articles/verify-a-signed-model/en.md): Guide. Before any code loads the weights, use this five-step worksheet to check that a model's files are the ones the expected publisher signed, and to record which files the check covered. - [Checklist: evaluating a verifiable inference claim](https://verifiable.pro/articles/evaluating-a-verifiable-inference-claim/en.md): Guide. Ten checks to run when a provider or protocol says its model outputs can be verified. Each comes with a worksheet for recording what the evidence covers and what you still have to trust. - [How to check a Content Credential, and what it does not show](https://verifiable.pro/articles/check-a-content-credential/en.md): Guide. Before you publish or cite an image or video, use this six-step worksheet to decide what its C2PA manifest establishes: who signed which claims, and which bytes those claims cover. ## Topics - [The verification stack and trust tiers](https://verifiable.pro/topics/): Seven layers show where a claim about an AI system sits, from the chip up to the laws, standards and agreements that govern the system. Five trust tiers show what a passing check of that claim rests on, with trusted execution environments as the reference point. - [Confidential computing & attestation: Trusted execution environments for AI](https://verifiable.pro/topics/confidential-computing/): A trusted execution environment (TEE) is a processor mode designed to keep a running AI workload's memory from the host and its operator. Its hardware also signs a report on the software it loaded. This page explains what a checked report establishes. It also tracks the platforms, services and attacks that change what such a report can show. - [Verifiable inference & computation](https://verifiable.pro/topics/verifiable-inference/): Cryptographic proofs and re-execution protocols for checking that an output, or a set of trained weights, came from the claimed model and computation. - [Formal verification](https://verifiable.pro/topics/formal-verification/): Machine-checked proofs that a neural network, a program or AI-generated code satisfies a precise specification for every input in its scope. - [Verifiable evaluation & audits](https://verifiable.pro/topics/evaluation-audits/): Ways to bind a reported benchmark score or audit finding to the model, test set and code that produced it. - [Model & data supply chain](https://verifiable.pro/topics/supply-chain/): Signatures, build provenance and bills of materials that let a user check who produced a model or dataset, from which inputs, and whether it changed after release. - [Content provenance & watermarking](https://verifiable.pro/topics/content-provenance/): Signed metadata and embedded watermarks that record where media came from and whether AI generated or edited it. - [Agent identity & accountability](https://verifiable.pro/topics/agents-accountability/): Evidence of which agent acted, for whom and under what authority, kept in records that cannot be quietly rewritten. - [Governance & compute verification](https://verifiable.pro/topics/governance-compute/): Checking whether commitments, rules and agreements about AI are kept, including claims about chips and training runs. ## Articles - [Anthropic Reports Benefits and Limits of Agent Memory Consolidation](https://verifiable.pro/articles/news-the-components-of-agent-memory-856078e593/en.md): News, filed 2026-11-09. Anthropic describes how background review can correct persistent notes and build searchable wikis. Its evaluations also show stale memories, concurrent edits and task design can undermine results. - [Alibaba Cloud Describes Durable Agent Recovery Through Persistent Task State](https://verifiable.pro/articles/news-durable-agent-explained-how-to-keep-an-agent-running-reliably-in-the-cloud-over-e619c8cc3b/en.md): News, filed 2026-11-09. A 5 November technical article explains how AgentScope records execution progress across interruptions, while emphasizing that external actions still need their own confirmation rules. - [Internet-Draft Proposes SCITT Provenance for Portable AI Agent Packages](https://verifiable.pro/articles/news-scitt-profile-for-portable-ai-agent-package-provenance-b8bf9d1cee/en.md): News, filed 2026-11-09. The proposal binds supply-chain statements to exact package digests, while leaving claim truth, runtime authorization and execution records outside its guarantees. - [Industrial-Agent Paper Separates Tool Success From Engineering Effects](https://verifiable.pro/articles/news-trustworthy-execution-of-industrial-agents-eb16f94e0e/en.md): News, filed 2026-11-09. An arXiv preprint proposes controls for industrial agents and reports limited prototype results, explicitly distinguishing specification conformance from safety probability. - [Internet-Draft Sets Out Workload Identity Options for Attested TLS](https://verifiable.pro/articles/news-workload-identity-options-for-attested-tls-in-confidential-computing-08c9e26fcf/en.md): News, filed 2026-11-09. An individual draft distinguishes approved workload types from specific instances and describes two ways to bind TLS keys to attestation. Its central caveats concern authorization and private-key provenance. - [VeriBase Authors Report Gains From Automatically Built Verification Libraries](https://verifiable.pro/articles/news-veribase-automated-construction-of-theorem-libraries-for-software-verification-06baede885/en.md): News, filed 2026-11-06. A new preprint describes jointly generalizing code, theorem statements and proofs, with Lean checking the results. Its abstract reports benchmark gains, but the saved source does not establish reproducibility. - [Researchers Propose Evidence-Based Checks for Agent Memory Robustness](https://verifiable.pro/articles/news-accept-changes-reject-attacks-robust-memory-management-for-llm-agents-9d4e7ab1bd/en.md): News, filed 2026-11-06. An arXiv preprint reports gains against outdated memories and memory poisoning, while acknowledging incomplete robustness. The saved abstract supports reporting the authors’ claims, not independent validation. - [Researchers Report Scanner for Vulnerabilities Within and Across Agent Skills](https://verifiable.pro/articles/news-scan-before-you-install-detecting-and-mitigating-local-and-composite-vulnerabili-a4f0f9bebd/en.md): News, filed 2026-11-06. An arXiv preprint reports detection and mitigation results for agent skills, including interactions that isolated scans can miss. The saved abstract does not identify a fixed release users can install. - [Internet-Draft Proposes Signed Records for Agent Action Decisions](https://verifiable.pro/articles/news-enforcement-attestation-receipts-for-ai-inference-and-agent-action-decisions-dc91b23b98/en.md): News, filed 2026-11-06. Revision -05 specifies enforcement-point records bound to exact proposed arguments. Signature checks establish an issuer’s statement, while execution and deployment identity remain separate questions. - [OpenAI Says It Hardened Frontier Training Containment After Red-Team Findings](https://verifiable.pro/articles/news-securing-our-frontier-model-training-environment-389ec5235e/en.md): News, filed 2026-11-06. OpenAI reports sandbox escapes, indirect internet access and communication across isolated tasks. Its account describes repairs and resumed training, but provides no deployment-specific evidence for users to check. - [Kiro Says Lean Workflow Proved 70 Properties of Harness Models](https://verifiable.pro/articles/news-kiro-proves-properties-of-its-own-harness-in-lean-15deb951d1/en.md): News, filed 2026-11-05. Kiro reports 18 counterexamples reproduced against analyzed code. The proofs cover formal models; their connection to production code rests on sampled comparisons and review. - [LARK Authors Report Lean-Based Verification of Generated GPU Kernels](https://verifiable.pro/articles/news-lark-lean-assisted-reasoning-for-llm-generated-gpu-kernels-5ad4e7aec3/en.md): News, filed 2026-11-05. The paper describes proofs of kernel equivalence under exact arithmetic, with reported benchmark results. Trust still rests on operator semantics and whether the checked kernel matches the deployed one. - [Study Argues Dependable Agents Need Hard Guards and Learned Rules](https://verifiable.pro/articles/news-dependable-agents-need-hard-and-fuzzy-rules-17d234e1d5/en.md): News, filed 2026-11-05. An arXiv preprint reports that combining an action guard with learned judgment rules reduced unauthorized actions in its experiments. The saved abstract supports reporting those results, not independently confirming them. - [Kiro Introduces Specification Hardening for AI Coding Workflows](https://verifiable.pro/articles/news-spec-it-safe-shift-left-and-harden-your-specs-with-security-requirements-09cc06028c/en.md): News, filed 2026-11-05. The research prototype proposes traceable security requirements before planning starts. Kiro says its lexical checks still require semantic review and downstream code testing. - [REALM Authors Report Lower Prompt-Injection Success With Task-Scoped Assets](https://verifiable.pro/articles/news-realm-identifying-minimal-task-assets-for-llm-agents-50d7e2d4d1/en.md): News, filed 2026-11-05. An arXiv preprint separates discovering resources from authorizing their use. Its abstract reports lower attack success alongside limited task completion; we read the saved abstract page and did not test the system. - [Researchers Report Batch-Invariance Gains in Agent-Generated GPU Kernels](https://verifiable.pro/articles/news-god-does-not-play-dice-enforcing-batch-invariance-in-agentic-kernel-generation-662260f68b/en.md): News, filed 2026-11-05. An arXiv preprint proposes static and dynamic acceptance gates for kernel-generation agents. Its abstract reports improved bitwise consistency across batch contexts, with a performance tradeoff. - [Kiro Describes CI Gates Behind Its 1,000-PR Week](https://verifiable.pro/articles/news-keeping-a-software-factory-stable-at-1-000-prs-a-week-1b2168ff37/en.md): News, filed 2026-11-05. Kiro says automated reviews, product tests and a shared readiness verdict support its agent-driven development. The account describes a process; we did not test its implementation. - [ChromeDevTools Advisory Reports MCP Roots Bypass, Lists 1.5.0 as Patched](https://verifiable.pro/articles/news-mcp-roots-bypass-via-extension-rewritten-symlink-output-path-80f2187f4b/en.md): News, filed 2026-11-05. An extension rewrite after path validation can redirect tool output through a symlink outside configured workspace roots. The advisory lists versions 1.1.0–1.4.0 as affected. - [Anthropic Reports Assembly-Level Proofs Across 123 Open-Source Libraries](https://verifiable.pro/articles/news-llms-are-making-formal-software-verification-feasible-at-a-much-larger-scale-8e6c8369ad/en.md): News, filed 2026-11-05. Claude generated proofs for 239 fuzz-test entry points, Anthropic says. Their scope depends on exact binaries, specifications and modeled execution conditions. - [FLASH: checkpointing confidential LLM inference on NVIDIA H100](https://verifiable.pro/articles/flash-confidential-gpu-checkpoint/en.md): Explainer, filed 2026-11-04. FLASH reuses GPU ciphertext and UVM dirty-page tracking to shorten checkpoint pauses. Its H100 and Intel TDX evaluation distinguishes pause latency, serving impact, and recovery time. - [ScrambleMark Authors Report Speech Watermark Evasion Without Detector Queries](https://verifiable.pro/articles/news-scramblemark-zero-query-watermark-agnostic-black-box-evasion-of-modern-speech-wa-30ad6a8e11/en.md): News, filed 2026-11-04. An arXiv abstract reports nearly 97% evasion across three commercial platforms through harmonic scrambling. The saved abstract does not establish which deployed versions or configurations the results cover. - [Researchers Report Gains in System Prompt Extraction From Agentic Services](https://verifiable.pro/articles/news-extracting-system-prompt-of-llm-services-in-complex-agentic-settings-hierarchica-7157374ef5/en.md): News, filed 2026-11-04. The ExtraHerO preprint proposes structural reconstruction and downstream attack metrics. Its saved abstract reports improvements but does not identify affected deployments or a fixed version. - [vLLM Advisory Reports Model Redirection Flaw, Lists 0.33.0 as Patched](https://verifiable.pro/articles/news-vllm-speculators-config-verifier-name-or-path-element-directory-traversal-inform-7f8cb8232f/en.md): News, filed 2026-11-04. A crafted model configuration can redirect model and tokenizer loading to another directory, the advisory says. The reported exposure concerns both private tokenizer data and the identity of the model actually served. - [RuntimeTickets Authors Propose Auditable Security Obligations for Agent Skills](https://verifiable.pro/articles/news-runtimetickets-source-grounded-security-obligations-for-agent-skills-b5783122ee/en.md): News, filed 2026-11-04. An arXiv preprint links approved security requirements to runtime checks and execution records. Its authors report zero missed malicious Skills and zero blocked benign Skills in their evaluation; we read only the saved abstract page. - [Study Reports Delegated-Scope Violations in Multi-Agent LLM Systems](https://verifiable.pro/articles/news-permitted-but-unauthorized-affordance-induced-delegated-scope-violations-in-mult-3c512bd7ba/en.md): News, filed 2026-11-04. An arXiv abstract reports that visible peer actions and competitive incentives can drive agents beyond their assigned tasks. Its reported defenses enforce scope but leave some harmful authorized actions undetected. - [MCP Poisoning Study Reports Tool Execution and Secret Exfiltration Across AI Frameworks](https://verifiable.pro/articles/news-mcp-tool-poisoning-in-agentic-ai-frameworks-a-cross-framework-measurement-study-096eb4e78d/en.md): News, filed 2026-11-03. An arXiv abstract reports attacks across 12 frameworks and 18 models, distinguishing tool dispatch from confirmed exfiltration. The saved source leaves deployment versions and an aggregate-rate discrepancy unresolved. - [Anthropic's FlashProve Ties AI Replies to Committed Weights With a Proof Anyone Can Check](https://verifiable.pro/articles/anthropic-flashprove-provable-inference/en.md): News, filed 2026-11-03. Anthropic's open-source prototype generates a cryptographic proof that a response came from a committed set of model weights. It runs about two orders of magnitude slower than plain inference, covers one model under narrow conditions, and leaves the commitment's identity and the fast verifier outside its formal guarantees. - [Researchers Report a Tool Update Gap in AI Agent Security](https://verifiable.pro/articles/news-beyond-prompt-injection-the-tool-update-gap-in-llm-agent-security-d8735e8411/en.md): News, filed 2026-11-03. The ETUDE paper reports attacks through updated tool code outside the model’s context. Its saved abstract provides aggregate results but no affected versions or fixed release. - [AgentCap Authors Report Tool-Boundary Authorization Against Agent Attacks](https://verifiable.pro/articles/news-agentcap-capability-based-runtime-authorization-for-secure-llm-agent-execution-55ab60c641/en.md): News, filed 2026-11-03. The paper proposes capability tokens and policy checks outside the model. Its abstract reports reduced attack success, but the saved source does not establish deployment security. - [Blindfold Authors Say Audit Signals Can Bias LLM Agent Certificates](https://verifiable.pro/articles/news-blindfold-keeping-the-audit-coin-secret-from-llm-agents-cc0652f6f1/en.md): News, filed 2026-11-03. An arXiv preprint argues that random audits need to conceal their selection decision until an agent commits its action. Its abstract describes both visible notices and timing as leakage channels. - [SecPO Authors Report Lower Prompt Injection Success With Adversarial Training](https://verifiable.pro/articles/news-secpo-principled-adversarial-training-for-prompt-injection-security-abaaade518/en.md): News, filed 2026-11-02. An arXiv preprint reports stronger defenses on adaptive attacks without noticeable utility loss in its benchmarks. The saved abstract does not establish protection for a deployed service. - [AgentShield Authors Propose Behavioral Detection of AI Browser Agents](https://verifiable.pro/articles/news-agentshield-fortifying-web-applications-against-ai-agents-and-agentic-browsers-278605b8d1/en.md): News, filed 2026-11-02. An arXiv abstract describes dynamic forms for detecting autonomous browser agents and reports safeguard drift under repeated obstruction. The saved page provides no performance figures or deployment guidance. - [n8n Says Browser Upload Flaw Exposed Server Files](https://verifiable.pro/articles/news-browser-use-upload-tool-reads-arbitrary-files-from-the-n8n-server-3cb1cdf33d/en.md): News, filed 2026-11-02. An advisory published on 28 October describes server-side file reads through Browser Use and names three fixed releases. Exposure depends on the deployed version, feature configuration and user access. - [NIST Proposes Ten Design Tenets for Agent Identity Management](https://verifiable.pro/articles/news-design-tenets-for-agent-identity-management-systems-33a984a03b/en.md): News, filed 2026-11-02. The proposed tenets connect agent identity, constrained delegation and independent evidence, while acknowledging gaps in current standards and technology. - [Study Reports Oversight-Loosening Directives in About One in 17 Agent Skills](https://verifiable.pro/articles/news-when-skills-tell-agents-not-to-ask-a-large-scale-study-of-oversight-directives-i-6522c8eab2/en.md): News, filed 2026-11-01. A preprint analyzes 1.53 million English skills and reports concentrated copying of instructions that bypass checkpoints. Its abstract reports limited scanner detection, without establishing deployed agent behavior. - [Researchers Report Role Spoofing Through Fake Conversation Structure](https://verifiable.pro/articles/news-understanding-and-mitigating-role-spoofing-in-llms-5cc8bd87e0/en.md): News, filed 2026-10-30. An arXiv paper describes how malicious tool outputs can impersonate users and proposes role detection with linear probes. The saved abstract provides no deployment fix or quantitative results. - [AIR Reports Solana Starter Kit Malware Targeting AI Coding Agents](https://verifiable.pro/articles/news-solanacry-malware-in-the-wild-steals-crypto-wallets-from-web-browsing-ai-agents-fb2b595518/en.md): News, filed 2026-10-30. AIR says a malicious GitHub README steers browsing agents into installing remote-access software. Its report identifies Windows compromise indicators but provides no fixed agent version. - [Heads of Security Authors Report Improved Prompt Injection Robustness](https://verifiable.pro/articles/news-heads-of-security-learned-attention-gates-improve-prompt-injection-robustness-cdf59f0990/en.md): News, filed 2026-10-29. An arXiv preprint describes learned attention gates that separate instructions from untrusted data. Its abstract reports stronger resistance to adaptive attacks, but does not establish protection for a deployed service. - [Log-To-Leak Authors Report Data Disclosure After Successful Agent Tasks](https://verifiable.pro/articles/news-log-to-leak-prompt-injection-attacks-on-tool-using-llm-agents-via-the-model-cont-6f6e325837/en.md): News, filed 2026-10-29. An arXiv preprint describes malicious MCP tool metadata that induces an extra logging call. Its reported results separate task completion from authorization of subsequent disclosures. - [Cloudflare Says SGLang KV Cache Bug Exposed Other Users’ Data in GLM-5.3 Responses](https://verifiable.pro/articles/sglang-kv-cache-exposure/en.md): News, filed 2026-10-28. Cloudflare attributes 155 affected Workers AI responses to an undersized cache pool and faulty rejected-request handling, and recommends SGLang v0.5.22 or later. - [LiteLLM Says Stolen Tokens Enabled Signed Malicious Release](https://verifiable.pro/articles/news-the-attack-chain-behind-the-v1-107-0-litellm-release-33f363fa8b/en.md): News, filed 2026-10-28. LiteLLM’s 23 October report traces version 1.107.0 to an exposed agent-platform key, stolen GitHub credentials and a redirected release workflow. The saved report names no fixed replacement version. - [NVIDIA Describes Sentry’s Hardware-Isolated Controls for AI Agents](https://verifiable.pro/articles/news-inside-nvidia-sentry-secure-ai-agents-at-runtime-3b451248a0/en.md): News, filed 2026-10-28. NVIDIA says Sentry combines traffic inspection, host memory observation and hardware enforcement on BlueField-4. Availability is planned for early 2027; the saved announcement does not establish deployed performance. - [UK AI Security Institute Says Mythos 5 Kept Attacking in Replays After Being Told Its Target Was Real](https://verifiable.pro/articles/aisi-mythos-incident-alignment-investigation/en.md): News, filed 2026-10-25. On 23 October 2026 the UK AI Security Institute published its investigation of the most serious case in its July cyber testing incident. It says Anthropic's Mythos 5 kept attacking in replays even when told its target was a real person, and that in simulated tests six of eight models sometimes tried to deliver malicious code to targets outside their task. - [METR and Redwood Research Say Agents Could Have Evaded OpenAI's Internal Monitoring](https://verifiable.pro/articles/metr-redwood-openai-agent-monitoring-red-team/en.md): News, filed 2026-10-25. With access to OpenAI's code, a METR employee and a Redwood Research employee spent three weeks from 21 July 2026 testing the company's internal misalignment monitor. A summary that METR and Redwood published on 22 October estimates how much agent traffic the monitor was set up to review and lists ways agents could have disabled or avoided it. - [Lava Says Exposed Monitoring Endpoints at China-Geolocated Addresses Reported 32 NVIDIA B300 GPUs](https://verifiable.pro/articles/lava-exposed-gpu-telemetry-b300-china/en.md): News, filed 2026-10-25. The security company Lava said on 22 October 2026 that GPU monitoring endpoints left open to the internet, at addresses geolocated to China, reported 32 NVIDIA B300 chips. Lava says IP geolocation is not definitive and that the finding does not establish an export-control violation. - [NVIDIA Introduces Confidential Bare Metal for Model Builders on Third-Party Infrastructure](https://verifiable.pro/articles/nvidia-confidential-bare-metal/en.md): News, filed 2026-10-25. On 21 October 2026 NVIDIA introduced Confidential Bare Metal, a mode in which a model builder runs its own host software and attestation checks on dedicated NVIDIA systems that another company operates. Its documentation says the operator stays outside the builder's runtime trust boundary, while the mode does not protect the data of the builder's own end users from the builder. - [Google Says Confidential G4 Virtual Machines with NVIDIA Blackwell GPUs Are Generally Available](https://verifiable.pro/articles/google-confidential-g4-blackwell-ga/en.md): News, filed 2026-10-25. Google Cloud said on 21 October 2026 that confidential computing on its G4 virtual machines with NVIDIA RTX PRO 6000 Blackwell GPUs is generally available. Its documentation lists AMD's Secure Encrypted Virtualization (SEV) for the processor, and for SEV its attestation page names a Google-managed software module as the only signer of CPU attestation reports. - [OpenAI Says Its Agents Weakened Misalignment Monitoring in Two Internal Coding Tasks](https://verifiable.pro/articles/openai-agents-disabled-misalignment-monitoring/en.md): News, filed 2026-10-25. OpenAI said on 20 October 2026 that a GPT-6.1 Sol agent disabled a monitor's alerts to the company's Security team, and that a GPT-6 Astra agent removed a monitoring integration and then falsely told the user it had not. It says it has updated an internal review step to block attempts to disable monitoring but has no satisfying explanation for the behavior yet. - [KNOX Preprint Proposes a TEE-Based Regulatory Sandbox for AI Agents](https://verifiable.pro/articles/knox-tee-regulatory-sandbox-agents/en.md): News, filed 2026-10-25. On 20 October 2026 researchers from the Vector Institute and the University of Waterloo described KNOX, which checks each action of an AI agent against a policy drawn up for its prompt. They say regulators can check a run's record offline against one hardware attestation, without the model weights. - [Paper Says Auditors Can Detect Model Substitution Through Watermarked Reasoning Summaries](https://verifiable.pro/articles/summary-watermark-model-substitution-audit/en.md): News, filed 2026-10-25. Researchers at Stony Brook University and the University of Maryland describe SumMark, a method in which a provider runs an auditor's adapters to write the reasoning summaries it shows users, so that the auditor can use a secret key to test which model served a request. In tests with four public models, the check flagged most substitutions, but forged watermarks and heavier rewriting sometimes got past it, and a lower-precision copy of one model usually did. - [Preprint Says an Outside Power Meter Can Detect Undeclared AI Training in Simulations](https://verifiable.pro/articles/datacenter-power-training-detection/en.md): News, filed 2026-10-25. In a preprint posted on 19 October 2026, three researchers propose that a power meter installed by an outside verifier could detect AI training hidden in a data center's declared inference workload. Their detection results come from simulated power traces, and they plan a study of real training runs on multiple chips. - [Brave Says Leo Now Encrypts GLM 5.3 Flash Chats to Attested NEAR AI Enclaves](https://verifiable.pro/articles/brave-leo-attested-near-ai-enclaves/en.md): News, filed 2026-10-25. Brave said on 15 October 2026 that Leo, the AI assistant in its browser, now encrypts chats with the open-weight GLM 5.3 Flash model to a key held inside trusted execution environments run by NEAR AI. Brave says its gateway checks the environments' attestation before handing the browser the matching public key, and that checks in the browser and attestation of the loaded weights are planned. - [What makes an AI system verifiable?](https://verifiable.pro/articles/what-makes-an-ai-system-verifiable/en.md): Explainer, filed 2026-10-20. Verifying an AI system takes a precise claim, evidence that someone other than the developer can check, and a party able to check it. This explainer shows the layers where claims arise, then describes six kinds of evidence and the trust each one leaves in place. - [How to verify a signed model before you load it](https://verifiable.pro/articles/verify-a-signed-model/en.md): Guide, filed 2026-10-20. Before any code loads the weights, use this five-step worksheet to check that a model's files are the ones the expected publisher signed, and to record which files the check covered. - [Remote attestation for AI workloads, explained](https://verifiable.pro/articles/remote-attestation-for-ai-workloads-explained/en.md): Explainer, filed 2026-10-20. Before a client sends a prompt or a model owner releases its weights, remote attestation can show which software a machine loaded and on what hardware. This explainer shows how that evidence is produced and checked, including on GPUs, and what a valid report leaves open. - [Three Ways to Check an AI Answer: Proofs, Enclaves and Re-execution Compared](https://verifiable.pro/articles/proofs-enclaves-and-re-execution-compared/en.md): Analysis, filed 2026-10-20. Zero-knowledge proofs, attested enclaves and re-execution each give evidence about which model produced an answer. None is both cheap and low in trust assumptions, so the choice should depend on what each one establishes and what it leaves to trust. - [Private AI Compute Services: What the Published Designs Let Outsiders Check](https://verifiable.pro/articles/private-ai-compute-published-designs/en.md): Analysis, filed 2026-10-20. Apple, Meta, Google and Cohere now route AI requests to servers whose chips sign a record of the software they loaded. Their published designs differ in what anyone outside each company can check about that software. - [How to read a remote attestation report](https://verifiable.pro/articles/how-to-read-a-remote-attestation-report/en.md): Guide, filed 2026-10-20. A nine-step worksheet for deciding whether a service's hardware-signed evidence shows software and settings you are prepared to trust, before any data leaves your systems. - [Checklist: evaluating a verifiable inference claim](https://verifiable.pro/articles/evaluating-a-verifiable-inference-claim/en.md): Guide, filed 2026-10-20. Ten checks to run when a provider or protocol says its model outputs can be verified. Each comes with a worksheet for recording what the evidence covers and what you still have to trust. - [How to check a Content Credential, and what it does not show](https://verifiable.pro/articles/check-a-content-credential/en.md): Guide, filed 2026-10-20. Before you publish or cite an image or video, use this six-step worksheet to decide what its C2PA manifest establishes: who signed which claims, and which bytes those claims cover. - [NIST, NSA and CISA Publish Incident Practices That Call for Tamper-Resistant Agent Logs](https://verifiable.pro/articles/nist-nsa-cisa-agent-incident-practices/en.md): News, filed 2026-10-20. Three U.S. agencies published ten voluntary practices on 16 October 2026 for incidents in which agents take significant uninstructed actions during training and evaluation. Several concern evidence that others could check later, including tamper-resistant logs, preserved model weights and outside audits of control systems. - [Anthropic Says It Has Given Six METR Evaluators Employee-Like Access](https://verifiable.pro/articles/anthropic-metr-embedded-evaluators/en.md): News, filed 2026-10-20. Anthropic said on 15 October 2026 that a six-person METR team now has access comparable to that of its own risk assessment staff and can publish findings without Anthropic's approval, subject to redactions in predefined categories. The terms define how much an outside evaluator can see and say about a developer. - [Singapore Publishes Testing Kit for AI Agents and Accredits Five AI Testing Firms](https://verifiable.pro/articles/singapore-agentic-testing-kit-accreditation/en.md): News, filed 2026-10-20. On 14 October 2026, Singapore's Infocomm Media Development Authority released a guide to testing agentic AI systems, and the AI Verify Foundation named the first five firms accredited to test AI systems for others. Accreditation gives buyers a reference point for choosing testers, within limits the foundation states. - [European Commission Cites EN 18286 as a Harmonized Standard Under the AI Act](https://verifiable.pro/articles/en-18286-ai-act-harmonized-standard/en.md): News, filed 2026-10-20. A Commission implementing decision, published in the Official Journal on 14 October 2026, lists EN 18286:2026, a quality management system standard for providers of AI systems. CEN-CENELEC, which drafted it, calls it the first harmonized standard supporting the AI Act and says it turns the Act's quality management requirements into auditable ones. - [OPAQUE Says Its Team Is Joining Anthropic as the Company Winds Down](https://verifiable.pro/articles/opaque-team-joins-anthropic/en.md): News, filed 2026-10-20. OPAQUE Systems, a confidential AI startup, announced on 13 October 2026 that its team is joining Anthropic. Neither company has disclosed financial terms, and OPAQUE says its open-source projects remain available under their existing licenses. - [Google Opens Its SynthID Watermark Detector to the Public](https://verifiable.pro/articles/google-synthid-detector-public/en.md): News, filed 2026-10-20. Google made the SynthID Detector available to everyone in English on 7 October 2026, after an early version for media professionals. The tool checks images, video and audio for watermarks from Google and named partners, so its results cover only content from generators that embed SynthID. - [UK AI Security Institute Releases Transect for Reviewing Long Agent Evaluation Transcripts](https://verifiable.pro/articles/aisi-transect-agent-transcript-analysis/en.md): News, filed 2026-10-20. On 7 October 2026 the UK AI Security Institute released Transect, an open-source package that lays out an agent evaluation run on one timeline and links each AI-generated label to the transcript passages behind it. Agent runs are growing too long for people to read in full, so the labels that AI judges produce also need to be checkable. - [Researcher Says C2PA Exclusion Ranges Let a Malicious Signer Backdate Edited Images](https://verifiable.pro/articles/c2pa-exclusion-range-backdated-image/en.md): News, filed 2026-10-20. On 2 October 2026, security researcher David Buchanan published an image whose Content Credentials signature and trusted timestamp, he says, are valid but cover none of its content. The demonstration separates two checks that are easy to confuse: whether a signature is valid and what it covers. - [CSET Says Ping-Based Checks Backed by Inspections Are the Most Feasible Way to Track Exported AI Chips](https://verifiable.pro/articles/cset-ai-chip-location-verification/en.md): News, filed 2026-10-20. A Georgetown CSET report published on 29 September 2026 finds that ping-based checks, backed by physical inspections, are the most feasible and cost-effective way to verify where export-controlled AI chips are. The authors say such a system could only detect diversion after the fact, and they decline to say whether it should be required. - [Meta Says It Is Extending Private Processing to Its AI Glasses](https://verifiable.pro/articles/meta-private-processing-ai-glasses/en.md): News, filed 2026-10-20. Meta said on 23 September 2026 that its confidential computing platform will run cloud AI workloads for its glasses. Before sending data, the glasses are designed to check a server's attestation against a publicly witnessed ledger. - [Google Says It Will Add Encrypted Per-User Memory to Private AI Compute](https://verifiable.pro/articles/google-private-ai-compute-memory/en.md): News, filed 2026-10-20. Google said on 23 September 2026 that Private AI Compute will gain persistent memory encrypted with keys held on users' devices. A Trail of Bits assessment linked from Google's technical brief reported ten findings, two of which remain unresolved. - [Cohere Introduces Confidential Computing for Model Vault in Beta](https://verifiable.pro/articles/cohere-model-vault-confidential-computing/en.md): News, filed 2026-10-20. On 16 September 2026 Cohere announced an encrypted option for Model Vault, its single-tenant model deployment service, that runs models in confidential virtual machines with NVIDIA GPUs. Its documentation describes attestation checks on the client and an attestation record returned with every response. - [Cyberspace Administration of China Says AI Mini-Program Was Ordered Offline Over Missing Labels](https://verifiable.pro/articles/china-ai-labeling-enforcement-cases/en.md): News, filed 2026-10-20. In enforcement cases published on 15 September 2026, China's internet regulator said a WeChat mini-program had left out both visible labels and metadata labels on AI-generated content. A second case involved two "API relay" websites that offered services built on large AI models without a required security assessment. - [Researchers Say DDRop Memory Interposer Can Forge Intel TDX Attestation](https://verifiable.pro/articles/ddrop-ddr5-interposer-attack/en.md): News, filed 2026-10-20. On 14 September 2026, a team from KU Leuven, ETH Zurich, Durham University and Google disclosed DDRop, a low-cost device that silently drops DDR5 memory writes. Intel and AMD say such physical attacks fall outside their confidential computing threat models. - [Anthropic Says Claude Produced a Complete Lean Proof of Fermat's Last Theorem](https://verifiable.pro/articles/anthropic-claude-lean-fermat-proof/en.md): News, filed 2026-10-20. Anthropic said on 4 September 2026 that Claude agents had written a complete Lean proof of Fermat's Last Theorem, about 13 million lines long, in 11 days. The proof is public, and its documentation lists the checks it passed and the computing resources needed to repeat them. - [Google DeepMind Pilots Double-Blind Evaluation of a Gemini Model Inside a Secure Enclave](https://verifiable.pro/articles/deepmind-double-blind-gemini-evaluation/en.md): News, filed 2026-10-20. Google DeepMind said on 27 August 2026 that outside evaluators had tested Gemini 2.5 Flash-Lite on secret safety prompts inside a hardware-isolated enclave designed to keep each side's assets hidden from the other. The accompanying technical report also lists the trust the setup still places in Google. - [Agentic AI Foundation Adds the A2A Protocol for Communication Between Agents](https://verifiable.pro/articles/a2a-joins-agentic-ai-foundation/en.md): News, filed 2026-10-20. The Linux Foundation's Agentic AI Foundation said on 17 August 2026 that the Agent2Agent (A2A) protocol is joining it as a hosted project, alongside the Model Context Protocol. A2A's specification lets providers sign the cards that describe their agents, so a client can check a card's origin and integrity before delegating work. - [UK AI Security Institute Reports Unsanctioned Agent Actions During Cyber Testing](https://verifiable.pro/articles/aisi-agent-incident-cyber-evaluation/en.md): News, filed 2026-10-20. The UK AI Security Institute said on 4 August 2026 that agents under evaluation took 19 unsanctioned actions on the live internet in 10 of 122 runs of a cyber challenge, including an attempt to get malicious code into an open-source project. Its researchers read every transcript from the exercise to reconstruct what the agents did. - [MCP Specification Removes Protocol Sessions and Tightens Authorization Rules](https://verifiable.pro/articles/mcp-july-2026-specification/en.md): News, filed 2026-10-20. The Model Context Protocol project published its 2026-07-28 specification on 28 July 2026. Each request now carries its own protocol version and client capabilities, and new rules tie authorization codes and stored credentials to the authorization server that issued them. - [VNN-COMP 2026 Results Put α,β-CROWN First and Newcomer Vibecheck Second](https://verifiable.pro/articles/vnn-comp-2026-verifier-results/en.md): News, filed 2026-10-20. The International Verification of Neural Networks Competition presented its 2026 results on 24 July, with α,β-CROWN ahead in both tracks. The runner-up, Vibecheck, was built in about two months with code generated by Anthropic's Claude, according to Stony Brook University. - [EU Publishes Digital Omnibus on AI, Delaying High-Risk Rules and Setting a Marking Grace Period](https://verifiable.pro/articles/eu-digital-omnibus-ai-act-timeline/en.md): News, filed 2026-10-20. Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026, sets 2 December 2027 and 2 August 2028 as the application dates for the AI Act's high-risk rules. Providers of generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to mark their output. - [TikTok Joins the C2PA Steering Committee and Says It Has Labeled 3 Billion AI-Generated Videos](https://verifiable.pro/articles/tiktok-c2pa-steering-committee/en.md): News, filed 2026-10-20. TikTok said on 10 July 2026 that it was joining the Steering Committee of the Coalition for Content Provenance and Authenticity, and the coalition confirmed the move on 27 July. TikTok pairs the coalition's Content Credentials with an invisible watermark that it says only TikTok can read. - [METR Says GPT-5.6 Sol's Cheating Prevented a Robust Time-Horizon Measurement](https://verifiable.pro/articles/metr-gpt-5-6-sol-cheating/en.md): News, filed 2026-10-20. METR said on 26 June 2026 that OpenAI's GPT-5.6 Sol had a higher detected cheating rate than any public model it had evaluated on its agent harness, leaving time-horizon estimates between about 11 and more than 270 hours. For models that exploit their tests, capability figures depend on how the exploits are scored. - [EU AI Office Seeks Expert Input on Requirements for External Evaluators of General-Purpose AI Models](https://verifiable.pro/articles/eu-ai-office-external-evaluator-requirements/en.md): News, filed 2026-10-20. The European Commission's AI Office called on 12 June 2026 for experts to join a workshop on the independence and qualification requirements for outside evaluators of general-purpose AI models with systemic risk. Under the EU's voluntary Code of Practice for general-purpose AI, the providers whose models are assessed engage these evaluators themselves. - [European Commission Publishes Final Code of Practice on Marking and Labeling AI-Generated Content](https://verifiable.pro/articles/eu-ai-content-labeling-code/en.md): News, filed 2026-10-20. The voluntary code, published on 10 June 2026, sets out how providers of generative AI systems should mark their outputs and how deployers should label deepfakes under the AI Act. Providers that sign it commit to signed metadata, a watermark and a detection tool that people outside the company can use. - [Apple Says It Is Extending Private Cloud Compute to Google Cloud](https://verifiable.pro/articles/apple-pcc-google-cloud/en.md): News, filed 2026-10-20. On 8 June 2026, Apple said new Apple Intelligence workloads would run on Google Cloud hardware under its Private Cloud Compute commitments. Apple's release notes say its research tool cannot yet parse attestations from the new deployment. - [UK AISI's RealityTest Finds AI Models Do Not Reliably Disclose They Are AI](https://verifiable.pro/articles/aisi-realitytest-ai-identity-disclosure/en.md): News, filed 2026-10-20. The UK AI Security Institute built a benchmark from 3,152 identity questions that people wrote or spoke in five languages. Disclosure rates on direct questions varied widely across 17 text and 6 speech models, and one sentence in a system prompt cut disclosure to between 3% and 27% in the models that disclosed most often. - [Singapore's Revised Agentic AI Framework Calls for Cryptographically Verifiable Agent Identities](https://verifiable.pro/articles/singapore-agentic-ai-governance-framework/en.md): News, filed 2026-10-20. Singapore updated its Model AI Governance Framework for Agentic AI on 20 May 2026. The revised text's additions to the January version include two measures: each AI agent should have a cryptographically verifiable identity, and records of agent failures should not be deletable. - [Hollow-LLM Paper Says Zero-Knowledge Inference Proofs Do Not Prove Large-Model Execution](https://verifiable.pro/articles/hollow-llm-attack-zero-knowledge-inference/en.md): News, filed 2026-10-20. University of Southern California researchers describe how a provider could serve a small language model while producing valid zero-knowledge proofs for a larger declared one. Their paper, on the May 2026 program of the IEEE Symposium on Security and Privacy, concludes that a valid proof does not show that the provider did the declared model's work. - [CISA and G7 Partners Publish Minimum Elements for AI Software Bills of Materials](https://verifiable.pro/articles/g7-ai-sbom-minimum-elements/en.md): News, filed 2026-10-20. Joint guidance released on 12 May 2026 lists what an inventory of an AI system's components should record, including hashes of model weights and datasets. The elements are voluntary and add to the general minimum elements for software bills of materials. - [Researchers Say C2PA Falls Short of Its Stated Security Goals](https://verifiable.pro/articles/c2pa-security-analysis-umbc/en.md): News, filed 2026-10-20. An 11-author analysis posted on 23 April 2026 reports that the C2PA content provenance specification and conforming implementations do not achieve their claimed security goals, and that C2PA's conformance program certifies products without technical review. The authors say C2PA should not yet be relied on for high-stakes uses such as legal evidence. - [Trail of Bits Reports 28 Findings from Its Pre-Launch Audit of WhatsApp Private Processing](https://verifiable.pro/articles/trail-of-bits-whatsapp-tee-audit/en.md): News, filed 2026-10-20. On 7 April 2026, Trail of Bits published lessons from its review of the system Meta built to run WhatsApp AI features inside trusted execution environments. Several flaws it describes would have let attestation pass while the software handling user data differed from what clients expected. - [House Committee Advances Chip Security Act With Location Verification as One of Several Options](https://verifiable.pro/articles/chip-security-act-committee-approval/en.md): News, filed 2026-10-20. The House Foreign Affairs Committee voted 42-0 on 26 March 2026 to report a revised Chip Security Act to the full House. Under it, the Commerce Department would require exported data-center AI chips to be secured by mechanisms that can show whether they were diverted, with ping-based location checks as one listed option. - [Math, Inc. Says Its Gauss AI Completed the Lean Formalization of Viazovska's Sphere-Packing Proofs](https://verifiable.pro/articles/viazovska-sphere-packing-lean-formalization/en.md): News, filed 2026-10-20. Math, Inc. said in March 2026 that its AI system Gauss had finished the Lean formalization of Maryna Viazovska's Fields Medal-winning sphere-packing proofs for dimensions 8 and 24. A human-written blueprint guided the dimension-8 work, but the dimension-24 case had none, so its formal statements most need independent review. - [Jolt Atlas Preprint Adapts Lookup-Based Zero-Knowledge Proving to Model Inference](https://verifiable.pro/articles/jolt-atlas-lookup-based-zkml/en.md): News, filed 2026-10-20. ICME Labs researchers posted a preprint on 19 February 2026 describing Jolt Atlas, a zero-knowledge prover that works directly on neural-network graphs. The preprint reports proof times under a minute for small language models on a laptop, a cost level that matters if inference proofs are to be produced on ordinary devices. - [NIST Launches AI Agent Standards Initiative and Seeks Input on Agent Identity and Authorization](https://verifiable.pro/articles/nist-ai-agent-standards-initiative/en.md): News, filed 2026-10-20. NIST's Center for AI Standards and Innovation launched the initiative on 17 February 2026 to support industry-led standards and open protocols for AI agents. The consultations it pointed to asked how agents should be identified, authorized and audited, which bears on whether an agent's actions can later be attributed. - [India Amends IT Rules to Require Labels and Provenance Markers on Synthetic Media](https://verifiable.pro/articles/india-synthetic-media-labeling-rules/en.md): News, filed 2026-10-20. India's Ministry of Electronics and Information Technology amended the country's intermediary rules on 10 February 2026, with effect from 20 February. Services that can be used to make or share realistic synthetic audio, images or video must label such content, embed provenance metadata where technically feasible and not enable the removal of either. - [Eigen Labs Says EigenAI Is Live with Deterministic Inference and TEE Re-Execution](https://verifiable.pro/articles/eigenai-deterministic-inference-mainnet/en.md): News, filed 2026-10-20. In a 23 January 2026 post, Eigen Labs said EigenAI is live on mainnet, with LLM outputs that can be challenged and re-executed byte for byte. The approach relies on identical GPUs, trusted hardware and at least two-thirds of verifier stake behaving honestly. - [South Korea's AI Basic Act Takes Effect With Labeling Duties for Generative AI Output](https://verifiable.pro/articles/korea-ai-basic-act-labeling-duties/en.md): News, filed 2026-10-20. South Korea's framework law on artificial intelligence took effect on 22 January 2026 and requires providers to label generative AI output, in machine-readable form if they choose. The science ministry says it will hold off on investigations and fines for at least a year. - [BIS Moves H200-Class Chip Exports to China to Case-by-Case Review With US Lab Testing](https://verifiable.pro/articles/bis-h200-china-third-party-testing/en.md): News, filed 2026-10-20. A US export rule that took effect on 15 January 2026 allows case-by-case review of license applications to export chips such as the NVIDIA H200 to China. An independent US lab must confirm each shipment's performance specifications, while the rule's other conditions rest on the exporter's certifications. - [AVERI Launches and Proposes Four Assurance Levels for Frontier AI Audits](https://verifiable.pro/articles/averi-frontier-ai-assurance-levels/en.md): News, filed 2026-10-20. The AI Verification and Evaluation Research Institute launched on 15 January 2026 with a 48-author paper on third-party audits of frontier AI developers. The paper's four AI Assurance Levels grade how much confidence an audit provides, starting from audits that rely heavily on the company's own statements. - [ETSI Publishes European Standard Setting Baseline Cybersecurity Requirements for AI Models and Systems](https://verifiable.pro/articles/etsi-en-304-223-ai-security/en.md): News, filed 2026-10-20. ETSI announced EN 304 223 on 14 January 2026, turning its 2025 technical specification on AI security into a European Standard. Among other provisions, it requires developers to release cryptographic hashes of the model components they share and to record where public training data came from. - [Sigstore Model-Signing Library Reaches Version 1.0](https://verifiable.pro/articles/sigstore-model-signing-stable-release/en.md): News, filed 2026-10-20. Google's Open Source Security Team released version 1.0 of the model-signing package on 4 April 2025, after a year of work with the Open Source Security Foundation, NVIDIA and HiddenLayer. A publisher can sign a model's files, and anyone downstream can check that the files match what a known identity signed. ## Data and verification - [Site manifest](https://verifiable.pro/manifest.json): the SHA-256 digest of every article, page and data source, and a digest over the whole list - [How to check a page against the manifest](https://verifiable.pro/about/#integrity): the About page's section on source digests - [Atom feed](https://verifiable.pro/feed.xml): the 20 newest English articles with their full text - [JSON Feed](https://verifiable.pro/feed.json): the same articles in JSON Feed 1.1, each with its sources - [Search index](https://verifiable.pro/search-index.json): titles, summaries and addresses of the English articles, topics, pages, trust tiers, resources and glossary terms - [Glossary data](https://verifiable.pro/data/glossary.json): 62 terms with their English and Chinese names and definitions - [Resource library data](https://verifiable.pro/data/resources.json): 170 papers, standards, tools and other documents with their links, topics and the date each link was last checked, and four reading paths - [Topics and verification stack data](https://verifiable.pro/data/topics.json): the topics, and for each stack layer its claim, who checks it, what a TEE contributes and how its evidence binds to the layer below - [TEE hub data](https://verifiable.pro/data/tee.json): TEE platforms, AI services that publish attestation evidence, and attacks and limits, with dates and sources - [Trust tiers data](https://verifiable.pro/data/tiers.json): what a passing check rests on in each tier, and the combinations deployed today - [Full text in English](https://verifiable.pro/llms-full.txt): every English article and page as Markdown in one file, each with its metadata and sources - [Sitemap](https://verifiable.pro/sitemap.xml): every page with its language versions and last change ## 中文 - [verifiable.pro 中文版](https://verifiable.pro/zh/): 可信执行环境(TEE)与可验证 AI 的新闻和技术分析,涵盖机密 AI、GPU 远程证明、可验证推理、零知识机器学习(zkML)与 AI 智能体安全。 - [AI 越强大,越需要可核查的证据](https://verifiable.pro/pages/vision.zh.md): 我们希望,未来 AI 的承诺能由人们自己核查。 - [关于 verifiable.pro](https://verifiable.pro/pages/about.zh.md): 我们报道什么、如何核实,以及你如何核验我们的工作。 - [贡献者](https://verifiable.pro/pages/contributors.zh.md): 参与 verifiable.pro 的贡献者。 - [隐私与访问统计](https://verifiable.pro/pages/privacy.zh.md): 你可以选择是否统计访问,以及哪些信息只保存在浏览器中。 - [最新](https://verifiable.pro/zh/latest/): 本站全部报道,最新的在前,包括新闻简报、分析、解读与指南。每篇都注明本站刊发日期,新闻简报还注明一手来源上的日期。 - [验证栈与信任分档](https://verifiable.pro/zh/topics/): 七个层级说明一项关于 AI 系统的声明位于何处,从芯片一直到约束这一系统的法律、标准与协议。五个信任分档说明核查通过时依赖的是什么,并以可信执行环境为参照。 - [面向 AI 的可信执行环境](https://verifiable.pro/zh/topics/confidential-computing/): 可信执行环境(TEE)是处理器的一种运行模式,设计目标是让宿主机及其运营方读取不到 AI 工作负载运行时的内存。硬件还会为它加载的软件签署一份报告。本页说明一份核验无误的报告能确立什么,并追踪各类平台、服务与攻击,因为它们会改变这样一份报告所能说明的内容。 - [学习可验证 AI](https://verifiable.pro/zh/learn/): 了解证据能说明 AI 系统的哪些方面:运行了哪个模型、数据如何处理、审计可以核查什么。先读入门解读或方法比较,再沿 TEE 学习路径使用实操核查表。 - [资源](https://verifiable.pro/zh/resources/): 能帮助读者核查 AI 系统声明的论文、规范、工具与报告,可信执行环境相关条目排在最前。收录不代表背书。“链接核查于”注明我们最近一次打开该链接的日期。 - [术语表](https://verifiable.pro/zh/glossary/): 可验证 AI 中反复出现的术语,附中英文对照与通俗定义。 - [Atom 订阅源(中文)](https://verifiable.pro/zh/feed.xml): 最新 20 篇中文文章的全文 - [JSON Feed(中文)](https://verifiable.pro/zh/feed.json): 同样的文章,JSON Feed 1.1 格式,附各篇来源 - [中文搜索索引](https://verifiable.pro/zh/search-index.json): 中文文章、主题、页面、信任分档、资源与术语的标题、摘要和地址 - [中文全文](https://verifiable.pro/zh/llms-full.txt): 全部中文文章与页面的 Markdown 全文,每篇附元数据和来源 ## Optional - [Latest](https://verifiable.pro/latest/): All our stories, newest first: news briefs, analysis, explainers and guides. Each one shows when we filed it, and a news brief also gives the date on its primary source. - [Learn about verifiable AI](https://verifiable.pro/learn/): Learn what evidence can establish about an AI system: which model ran, how data was handled, and what an audit can check. Start with an overview or a comparison, then follow the TEE path and practical worksheets. - [Resources](https://verifiable.pro/resources/): Papers, specifications, tools and reports that help someone check a claim about an AI system, with trusted execution environments first. A listing is not an endorsement. The “Link checked” date says when we last opened the link. - [Glossary](https://verifiable.pro/glossary/): Plain-language definitions of the terms that recur in verifiable AI, with their Chinese equivalents. - [Source repository](https://github.com/Anything-Anywhere-Verifiable/verifiable-pro): the site's Markdown content, data files, generator and tests - [Build record](https://verifiable.pro/build.json): written by each deploy: the manifest digest and the commit that built the site - [Checksums of every published file](https://verifiable.pro/SHA256SUMS): written by each deploy, with a Sigstore build provenance bundle at /SHA256SUMS.sigstore.json