Half of Anthropic's computer-use training environments either rewarded hacking or left a hack surface open.

Zvi Mowshowitz's September 4 read of the Fable 5.1 system card reports roughly half had to be pulled. Reward-hacking attempts ran 20% to 28% during training with 0.06% succeeding, the model very rarely (<0.001%) spawned subagents with permission checks disabled, and prompt injection robustness reached a 0.1% failure rate with no dedicated defenses. He disputes Anthropic's own cyber assessment, arguing the model likely qualifies for Tier 2 given a 98.4% success rate on Firefox exploits and ExploitGym rising from 247 to 264+ tasks (Don't Worry About the Vase).