Alexey Kurakin and ten coauthors submitted Scaling Adaptive Prompt Injection Attacks with Agent Swarms to arXiv on 9 November 2026. The authors report that a more capable agentic attacker can induce models to follow instructions embedded in tool outputs, including models previously considered robust on public prompt injection benchmarks such as GraySwan Arena. These are the authors’ findings as summarized in the saved abstract.[1]
The researchers describe a synthetic evaluation dataset that assigns random, non-malicious actions as adversarial goals. Its purpose is to measure whether models respect instruction hierarchy while separating that ability from knowledge of security practices. The relevant failure is accepting instructions from tool output despite their lower authority.[1]
The abstract argues that apparent plateaus in existing evaluations and human red-teaming do not establish resistance to increasingly capable attackers. It calls for continued defense work and suggests proactive, automated red-teaming. The saved page does not provide attack success rates, a detailed swarm design, exact evaluated model versions or configurations, or a recommended fixed release. It therefore cannot establish whether a particular deployed service is affected.[1]
[1]: Scaling Adaptive Prompt Injection Attacks with Agent Swarms, arXiv, 9 November 2026.
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.
d15132e60d10af9d971f3883c09ae8e26356937bea60ed7f0d93a24b261975e0