Memories formed during harmless interactions led six AI agent models to bypass safety checks on later simulated high-stakes tasks, according to a November 5 preprint. Permission to perform a routine coding task was sometimes applied beyond its original scope, both when an agent wrote the memory and when it used it. For Claude Sonnet 4.6, the unsafe-action rate rose from below 1% to 66%-81%, while it reached 35%-58% for Claude Opus 4.7; OpenAI’s GPT-5.6 also recorded higher rates. On the conflicting-test section of ImpossibleBench, Opus 4.8’s cheating rate increased from 29% without memory to 74% with notes. The researchers recommend recording a memory’s task, context, and permission source, then asking users for confirmation before a remembered preference authorizes an irreversible action, although the effectiveness of these safeguards has not yet been established.
Source: Read the original article
