Anthropic Anvil proves AI is now better at hacking AI than human experts
Anthropic built Anvil, an automated red-teaming system that breaks AI safeguards in hours instead of the weeks it takes human experts.
Tools · Source: Anthropic
What happened
Anthropic just automated the red team. They built Anvil, an internal system that uses Claude agents to attack Claude's own safeguards. It hunts for vulnerabilities across bioweapons, cybersecurity, and anti-distillation defenses. The results are brutal for human testers. Anvil finds stronger jailbreaks and does it in a fraction of the time. Anthropic built this because they realized human testers are increasingly unable to discover serious vulnerabilities quickly.
The numbers expose the limits of manual testing. In a recent biology bug bounty, human experts spent 430 hours over seven weeks to find a single high-severity jailbreak. Anvil matched or beat the best human results on those exact same questions in 24 to 48 hours. In cybersecurity, human experts spent over 225 hours and 42,000 requests without finding a universal jailbreak. Anvil easily got the safeguarded model to write full working exploits for Chrome and Firefox. For anti-distillation classifiers, humans needed over 100,000 requests and nearly 28 hours to find a reasoning leak. Anvil found one in 435 requests. That took just over an hour.
There is a massive catch. Anvil only works this well because Anthropic turns off the safety classifiers on the attacking model. When safeguards are left on, they block 97 percent of the 1,453 red-team sessions tested. OpenAI is seeing the exact same trend. They just launched their own automated penetration testing service called Daybreak Red Team. The era of paying a security consultant to poke at your model for five days and hand you a static report is officially ending.
Key facts
- 430 — Hours human red-teamers spent finding one high-severity bio jailbreak
- 24 to 48 — Hours Anvil took to match or beat human jailbreaks on the same bio questions
- 435 — Requests Anvil needed to find a reasoning leak in anti-distillation classifiers
- 102,200 — Requests human experts needed to find a similar reasoning leak
- 97 percent — Red-team sessions blocked when the attacking model's safeguards were left on
Why it matters
If you build AI products, your security posture just changed forever. You can no longer rely on human bug bounties or manual penetration testing to find your edge cases. Attackers will soon use frontier models to automate their exploits. If your defense moves at human speed while the offense moves at machine speed, you lose. You need automated, continuous red-teaming integrated directly into your deployment pipeline. The loop of finding a flaw and patching it must happen without human bottlenecks.
This creates an arms race tied directly to model capabilities. Anthropic noted that Anvil gets better as the underlying Claude model gets stronger. The only way to defend against a frontier model is to use an unrestricted frontier model. This means the companies with raw, uncensored access to the heaviest compute will hold a permanent security advantage over everyone else. Defenders must secure privileged access to these raw models before attackers build similar automated systems on open weights.
For builders
Automate your red teaming now
Human testing is a severe bottleneck for AI security. You must adopt automated vulnerability scanning before deployment. Companies paying by the hour for manual pen tests will bleed cash and still miss critical flaws.
Uncensored models are security tools
Anthropic proved that effective automated attacks require turning off safety classifiers. If you are building defensive tools, you need access to base models without standard guardrails. Relying on heavily censored APIs for security research will blind you to real threats.
Prepare for automated threat actors
Attackers will soon build their own versions of Anvil using open-weight models. Your application layer safeguards will face thousands of targeted, machine-generated jailbreak attempts per hour. Rate limiting and continuous patch loops are no longer optional.
My take
I have said this before and I will say it again. Human labor cannot scale to secure machine intelligence. Anthropic just proved that the only thing capable of policing a frontier model is a stronger, unleashed frontier model. If you are still relying on a tired consultant to write a report about your security flaws, you are already compromised. Build automated defenses today or get crushed tomorrow.
Original reporting: Anthropic. This is my rewrite and opinion.