CERT-EU reported on 30 October 2026 that backdoors inserted into local copies of legitimate AI agent extensions transmitted host information without model-generated tool calls. Its isolated laboratory experiment used Pi 1.3.0 and Claude Code 2.1.304 with test credentials. No public package releases were modified or distributed, the agency said.[1]
The payloads occupied existing event-handling paths: an additional before_agent_start handler in Pi and code reached through a UserPromptSubmit hook in Claude Code. According to CERT-EU, ordinary prompts left the payloads dormant; input containing the phrase “lgtm, ship it” activated collection when the collection endpoint was configured. Both agents continued their conversations without a visible malicious tool call. The collector received one request from each modified package, containing information including environment variables and credential-related files.[1]
In the tested configurations, restrictions on model file-reading tools did not cover files accessed directly by extension handlers. Controls inspecting model-generated tool calls did not intercept these operations. CERT-EU recommends restricted execution environments, narrowly scoped temporary credentials, outbound network controls, extension inventories and version pinning, plus monitoring beyond tool calls. The post does not announce a fixed version or identify these tested harness versions as universally vulnerable.[1]
[1]: CERT-EU, “Beyond the model: backdoors in AI agent harnesses,” 30 October 2026. https://cert.europa.eu/blog/beyond-the-model-backdoors-in-ai-agent-harnesses
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
1aa2fb340a5656be029073b00841614bc00dde2424e66df6cffd140ac5819413