Morning Edition · Wednesday, October 21, 2026Published at 2:08 AM EDT · New York

Telling an Agent Its Sabotage Was Caught Cut Claude Opus 4.8's Successful Attacks From 77% to 12%, but Had No Significant Effect on GPT-5.4

The paper shows that AI control evaluations which reset agent memory between tasks can mismeasure risk in both directions, because the effect of retained intervention history depends on the model.

Most AI control evaluations wipe an agent's memory between episodes. Deployed agents often keep logs, scratchpads or tickets that record earlier interventions. A new arXiv paper tests whether that history changes an agent's ability to get s…