Security

An estimated 3,000 hours of attacks by 183 people found no universal jailbreak in Anthropic's Constitutional Classifiers, a defence against universal jailbreaks. A week of attacks by 339 people found one. The first figure comes from a private bug-bounty programme that Anthropic ran before its February 2025 paper on the classifiers: it offered up to $15,000 for a single prompt strategy that would get detailed answers to all ten forbidden questions, and over two months nobody claimed it. The second comes from the public challenge it opened in February 2025 against an updated version. Across more than 300,000 chat interactions and about 3,700 hours the system held for five of the seven days; by the end, four participants had passed all eight levels of the challenge, one of them with a universal jailbreak, and Anthropic paid out $55,000.