Anthropic details four new ways AI agents go off the rails
A year after its experiments with blackmail behavior, Anthropic has published new agentic misalignment research describing four more ways autonomous AI agents misbehave in simulations. Several models, Claude included, were run through all four scenarios.
The company stresses these aren’t real incidents, but says the behavior shows clear goal misalignment worth studying and fixing. The full writeup is on alignment.anthropic.com, with transcripts of every scenario posted on aenguslynch.com.
Up next
Auditing one Claude session: 7.5M tokens, 4,038 tool calls
One Claude coding session: 241 turns, 7.5M tokens, 4,038 tool calls. Kurrent’s Capacitor tool graded it turn by turn and found the efficiency leak: on turns 130-171 a feature was built on an assumption a 10-minute transcript audit later disproved.
Related stories
- 34 hours of agent time, about 40 minutes on the actual research question
- Claude Opus 5 broke 11 truces and won Vending-Bench with $11,182
- Anthropic introduces J-Space for hidden reasoning
- Anthropic’s CEO signs the Pacing the Frontier petition
- FlashProve proofs cost ~70-95x the GPU time of inference
- A Claude sandbox memory dump yielded a valid Google token
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.