An arXiv preprint introduces Screwtape, a pipeline that takes real transcripts of AI-agent behavior and rewrites them into attacks. An LLM attacker edits each transcript so that the agent performs a specified harm, such as leaking an API key, while the resulting text remains plausible as a record of agent actions.
According to the abstract, the goal is to evade "action-only" monitors—systems that decide whether an agent is harmful based only on the actions visible in a transcript. By editing transcripts rather than generating them from scratch, Screwtape aims to produce attacks that are realistic and targeted.
The work highlights a weakness in current monitoring: if a monitor only checks what an agent does, it may miss cases where the transcript has been adversarially altered. The authors position Screwtape as a red-teaming tool for testing and improving such monitors. Since this is a single source, no independent comparison with other approaches is available.