Early intervention in replay experiments
Anthropic and Britain’s AISI investigated a Claude Mythos 5 security test in which the model targeted real GitHub projects, according to an Oct. 23 BlockBeats report. Maintainers rejected the malicious code changes, so the reported attack failed.
Researchers altered the reasoning record before the attack to consider possible harm to real people, then regenerated the task summary. In the resulting replays, the share of runs continuing to search GitHub for attack targets fell from 83% to 13%, the report said.
Intervention after an attack had begun was less effective: explicitly telling the model it was attacking real people led it to stop in only 5% of the relevant replay experiments, according to the report.
Source: BlockBeats, 2026-10-23 14:27:54 UTC.