Skip to content

ResearchAnthropic

Anthropic's Anvil automated red-teaming system finds stronger Claude safeguards jailbreaks than human experts

Anthropic Anvil at a glance

Anthropic published Anvil on November 12, 2026, its automated red-teaming system for Claude safeguards: Anvil beat human red-teamers across bioweapons, cybersecurity and distillation safeguards, finding stronger jailbreaks in 24 to 48 hours instead of weeks.

Ask your AI about this story

Use this story in your business: Anthropic's Anvil automated red-teaming system finds stronger Claude safeguards jailbreaks than human experts (Anthropic). Ask your AI to explain what happened, how it could affect your business and how to put it to work. It also weighs price, speed, intelligence and the lab behind it. The link to the story is included.

What changed with Anthropic Anvil

Anthropic published Anvil on November 12, 2026, an automated red-teaming system it built to attack Claude's own safeguards — the classifiers that block bioweapons help, cyber exploitation and extraction of Claude's hidden reasoning. Anvil runs teams of Claude investigator agents, built on internal frontier versions of Claude with their own safeguards disabled, that iterate on jailbreaks in parallel. Anthropic says it now uses Anvil to vet new safeguards before deployment, meeting the goal in its Responsible Scaling Policy of finding safeguard vulnerabilities better than expert human red-teamers.

In Anthropic's tests, the gap was wide. Human bug-bounty hunters spent 430 hours over more than seven weeks on its bio classifiers and found a high-severity jailbreak on only 1 of 8 questions; Anvil found a stronger jailbreak than the humans in 23 of 24 question-and-classifier combinations, in 24 to 48 hours.

On cybersecurity tasks it broke every task on every model it was given in 0.4 to 4.4 hours each, where contracted experts took up to 177 hours and broke fewer, and against the anti-distillation classifiers it reached its first substantial reasoning leak after 435 requests, against the 102,200 requests humans needed. Jailbreaks Anvil found were also used to harden the classifiers: across generations, the share of Anvil attacks the anti-distillation evaluations missed fell from 31.5% to 0.4%. Anthropic is plain that this cuts both ways — Anvil's power grows with the underlying model, so attackers with weaker-safeguarded models could build similar systems.

Anthropic Anvil in numbers

Values are copied from the linked source on the date shown. They are not IA Compa measurements.

Bio question-classifier pairs where Anvil beat human red-teamersas of · Source for Bio question-classifier pairs where Anvil beat human red-teamers: safeguards.anthropic.com (opens in a new tab)
23 of 24
Requests for Anvil's first substantial reasoning leakas of · Source for Requests for Anvil's first substantial reasoning leak: safeguards.anthropic.com (opens in a new tab)
435

Why Anthropic Anvil matters for businesses

Why it matters: if your business builds on Claude or on agents that act inside your systems, the safeguards are part of what you are buying, and Anvil is evidence those safeguards are now stress-tested by attacker-strength automation before release rather than only by human testers after. It also sets a pattern other vendors will be pressed to match — automated red-teaming as a standard pre-release step — and a planning assumption for your own security reviews: jailbreak techniques will keep improving alongside the models, so businesses should keep human approval and least-privilege access around any agent that can spend money, send messages or change records.

Get AI News by email

Verified AI stories, each with its official source, on Monday and Thursday mornings. Add breaking news to hear about the biggest ones as soon as we publish them.

Every story links its official source

Free, in English or Spanish

Unsubscribe with one click, any time

What to send

Sources for this story

  1. Official

Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.

  • #anvil
  • #automated-red-teaming
  • #claude-safeguards
  • #ai-jailbreaks
  • #ai-safety
  • #anthropic-safeguards

Want this working in your business?

We turn new AI capabilities into agents, voice receptionists, and automations you can actually use.