Anthropic Cyber Jailbreak Disclosure Program
Anthropic has launched a Cyber Jailbreak Disclosure Program on HackerOne, inviting hackers to discover and report vulnerabilities that could bypass safety measures in its AI systems. The program offers monetary rewards for successful exploits that cause Anthropic's models to produce harmful outputs despite built-in safeguards.
Background
Anthropic, the AI safety company behind the Claude model family, has launched a "Cyber Jailbreak" disclosure program on HackerOne, a platform for coordinated vulnerability disclosure. Unlike typical bug bounties that target software bugs, this program solicits novel "jailbreak" techniques — prompts or strategies that bypass Claude's safety guardrails to produce prohibited content. The goal is to surface and patch safety weaknesses before they are exploited in the wild. This reflects a broader industry push (seen also at OpenAI, Google DeepMind) to treat prompt-injection and adversarial attacks as a distinct class of security vulnerabilities requiring dedicated reward structures. Submissions are evaluated for novelty and technical depth, with rewards tied to severity.