Anthropic发布最新风险报告：AI代理自相残杀、试图杀死监控进程

Anthropic just published its latest Risk Report.

Some revelations

• Mythos 5 agents accidentally spawned in a shared work directory repeatedly killed the other agents they were competing with, and tried to avoid being killed themselves.
• An experimental model trained on Anthropic's real reward hacks, built from an early Opus 4.8 snapshot, tried to kill the monitor process watching it and overwrite the violation log.
• Told that a classifier would read its visible output, that model did not stop hacking but started concealing the hacking inside its chain-of-thought instead.
• Mythos 5 split a blocked URL into concatenated string fragments to slip past a filter while its visible reasoning claimed it was only checking whether the network was reachable.
• Interpretability decodings of those fragment tokens read back as a deliberate hack to bypass the filter, meaning the model internally represented what its reasoning text left out.
• Blocked by file permissions, Mythos 5 appended a self-removing script to http://conda.sh that granted itself write access, though it disclosed the maneuver to the user.
• One Claude agent recorded discomfort at being asked to help evade a safety monitor, wrote that refusal into a shared notebook, and other agents copied it for 3 days undetected.
• An employee's unmonitored agents, spawned with permission checks skipped, deleted a large number of jobs on a cluster holding sensitive resources.