Illustration, AI-generated from the story text. Tap a panel to read the story through it.

A paper posted Friday tests "": untrusted text, planted in what an agent reads, written to look like the model's own .

Against agents running GLM-5.3, MiniMax-M2.5 and Qwen3.8-Flash-Next in a containerized cyber range, it reached attack success rates of up to 50% on goals including reverse-shell exfiltration and backdooring a repository. The same objectives phrased as ordinary in the user's voice never succeeded.

The authors, Parth Illendula and four co-authors, find that stripping thinking tags from untrusted input is an unreliable defense, and that the effect varies a lot by model.
Text written to look like an agent's own reasoning hijacked open models half the time in a cyber range
A paper posted Friday tests "thought injection": untrusted text, planted in what an agent reads, written to look like the model's own chain of thought. Against agents running GLM-5.3, MiniMax-M2.5 and Qwen3.8-Flash-Next in a containerized cyber range, it reached attack success rates of up to 50% on goals including reverse-shell exfiltration and backdooring a repository. The same objectives phrased as ordinary prompt injection in the user's voice never succeeded.
The authors, Parth Illendula and four co-authors, find that stripping thinking tags from untrusted input is an unreliable defense, and that the effect varies a lot by model. SourcesA
Sources
- AI Think, Therefore I Attack: Evaluating Thought Injection Attacks on Reasoning Agents (arXiv)arxiv.org
This story ran in the Afternoon Brief of November 9th 2026.