Claude Jailbreak results are in, and the hackers won

Claude Jailbreak results are in, and the hackers won Ad Skip to content Claude Jailbreak results are in, and the hackers won Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner Feb 15, 2025 Anthropic Update – Feb 15, 2025 Jailbreak update 2 Update from February 15, 2025: The results from Anthropic's Claude jailbreaking challenge are in. After five intense days of probing - involving over 300,000 messages and what Anthropic estimates was 3,700 hours of collective effort - the AI system's defenses finally cracked. Jan Leike, an Anthropic researcher, shared on X that four participants successfully made it through all challenge levels. One participant managed to discover a universal jailbreak - essentially a master key to bypass Claude's safety guardrails. Anthropic is paying out a total of $55,000 to the winners. Ad The challenge demonstrated that safety classifiers, while helpful, aren't sufficient protection on their own, Leike says. This aligns with what we've been learning from other recent AI safety research - there's rarely a silver bullet solution, and the probabilistic nature of these models makes securing them particularly challenging. Ad DEC_D_Incontent-1 Leike emphasizes that as models become more capable, robustness against jailbreaking becomes a key safety requirement to prevent misuse related to chemical, biological, radiological, and nuclear risks. Update from February 11, 2025: Ad Within just six days of launching the challenge, someone managed to bypass all the security mechanisms designed to protect Anthropic's AI model. Jan Leike, a former OpenAI alignment team member now working at Anthropic, announced on X that one participant successfully broke through all eight levels of the challenge. The collective effort involved around 3,700 hours of testing and 300,000 messages from participants. However, Leike notes that no one has yet discovered a universal jailbreak that could solve all challenge levels at once. Ad DEC_D_Incontent-2 After ~300,000 messages and an estimated ~3,700 collective hours, someone broke through all 8 levels. Ad However, a universal jailbreak has yet to be found... https://t.co/xpj2hfGC6W — Jan Leike (@janleike) February 9, 2025 As AI models become more capable, protecting them becomes increasingly critical - and universal jailbreaks grow more valuable. This dynamic suggests that language models might eventually develop their own security ecosystem, similar to what exists today for operating systems. Original article from February 04, 2025 The AI company Anthropic has developed a method to protect language models from manipulation attempts. Anthropic has developed a new safety method called "Constitutional Classifiers" to prevent people from tricking AI models into giving harmful responses. The technology specifically targets universal jailbreaks - inputs designed to systematically bypass all safety measures. To put the system through its paces, Anthropic recruited 183 people to try breaking t

... (truncated, 121 KB total)