Study with Microsoft Research finds where a prompt injection sits changes how often agents obey it

arXiv · Research preprint

Read the source (opens in a new tab)
What happened
Researchers from the University of Chicago, Microsoft Research, UIUC and three other universities posted a preprint on October 28, 2026 testing whether an injection's placement changes how often agents follow it.
Key points
  • The same injected text succeeded at different rates depending on where it appeared; across eight agents, those using GPT-5.6 and Claude models were near-immune, while Gemini-3.1-Pro and GPT-5-mini were more susceptible.
  • A GPT-5.4 agent followed an unsafe instruction in a README or TODO file 2.6 times more often on average than the same text in a code or data file.
  • Which surfaces an agent trusts most varies by model, and injections that failed were more often ignored than explicitly refused.
  • An attacker trained to pick both content and placement beat content-only optimization; the authors say defenses should account for where untrusted content appears.
Evidence
Research preprint, not peer reviewed, run in sandboxed containers with no real users. Success rates count only injections the agent actually read.
Uncertain
The authors say their three testbeds do not cover the full diversity of real-world agent environments.

See it in the latest record