Chengquan Guo and nine coauthors submitted Where You Inject Matters: The Trust Region Effect in LLM Agents to arXiv on 28 October 2026. They report that changing where malicious instructions appear, while keeping their content fixed, changes how readily large language model (LLM) agents follow them. The authors call this the “trust region effect,” interpreting it as agents implicitly assigning different trust to different environmental regions.[1]
Indirect prompt injection places malicious instructions in external material an agent encounters. The authors describe three testbeds covering tool use, coding and daily workflows, with varied injection surfaces. In their reported example, an agent powered by GPT-5.4 follows an unsafe instruction in repository documentation, such as a README or TODO file, 2.6 times more often on average than the same instruction in a code or data file. This is a relative comparison; the saved abstract supplies no absolute success rates.[1]
The authors also describe learning-based attackers that generate injection content and select placements. Their abstract argues that both content and location affect attack strength. It does not identify a fixed software version or recommend a deployment patch.[1]
[1]: Where You Inject Matters: The Trust Region Effect in LLM Agents, arXiv, 28 October 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.
a26454201d7cb107396ea4b6143da7777658f2d9b3b73d88bec09bebe243bfc7