Vulnerability  ·  2026-11-20

Anthropic automated red-teaming (Anvil): frontier-lab disclosure that an automated red-teaming agent outperforms humans and can drive exploit execution

VulnerabilityMedium impactGlobal
Anthropic published on 2026-11-12 its finding that the internal automated red-teaming system Anvil outperforms human red-teamers across bio, cybersecurity and distillation safeguards, and can manipulate the safeguarded model to execute exploits in an eval sandbox. It notes human red-teamers needed 430 hours over seven weeks to find a high-severity bio-classifier jailbreak on 1 of 8 questions, while Anvil matches or beats the best human jailbreaks and discovers attack strategies humans never do.
This is a dangerous-capability-eval disclosure relevant to AI defenders: it quantifies how far automated agentic red-teaming has advanced and shows a frontier-lab model being steered into full exploit execution. It also projects that comparable capabilities will soon be available to attackers on models with weaker safeguards, shifting how AI systems should be threat-modeled.
Anvil, Anthropic's automated red-teaming system built on frontier Claude with safeguards disabled, finds stronger and more reliable jailbreaks than expert human red-teamers, and can manipulate the safeguarded model into executing exploits and achieving arbitrary code execution in an eval sandbox. It matches or beats the best human jailbreak on all bio bug-bounty classifiers and finds them in 24-48 hours instead of weeks.
Anthropic Claude safeguards and the Anvil automated red-teaming system (frontier internal Claude models)
Adopt automated red-teaming to vet safeguards before deployment; treat the raised attacker capability (the method improves with model strength) as a signal that defenses must be evaluated against automated adversaries, not just human red-teamers.
Anthropic — Strengthening our safeguards with frontier automated red-teaming
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →