Anthropic offers $15k bounties to crack its AI safeguards

AI red-teaming just got a lot more official, and a lot more lucrative.

Anthropic announced an expansion of its model safety bug bounty program in partnership with HackerOne. Instead of just poking at models that are already out in the wild, researchers are being invited to break Anthropic's next-generation safety mitigations before they ever hit production.

Hunting the universal jailbreak

Anthropic is offering bounties of up to $15,000 for novel vulnerabilities. But they aren't looking for standard prompt tricks that get Claude to swear or write a mild diss track.

The initiative is strictly laser-focused on what Anthropic calls universal jailbreaks. These are repeatable, broad exploits that reliably bypass safety guardrails across diverse topics rather than single edge cases.

Specifically, Anthropic wants to know if a jailbreak can consistently force the model to answer a defined set of restricted questions in catastrophic threat vectors. That means chemical, biological, radiological, and nuclear weapons, alongside dangerous cybersecurity exploits.

Participants will get early, controlled access to test defensive layers that haven't been publicly deployed yet. The program is invite-only for experienced security researchers, with applications closing on Friday, August 16, and selections rolling out in the fall.

Why it matters

Most commercial AI safety work to date has felt like an endless game of whack-a-mole. A rogue user on Reddit discovers a clever roleplay framing, the model misbehaves, the company pushes a quiet patch, and the cycle repeats 48 hours later.

Anthropic is trying to get ahead of that loop. Handing pre-deployment access to external hackers is a smart admission that in-house red teams simply cannot match the collective ingenuity of the broader security community.

We think the $15,000 cap is somewhat modest compared to traditional six-figure zero-day bounties in cybersecurity, but it sets a meaningful baseline for AI-specific vulnerability research. It also shows Anthropic treating safety architecture as an independent engineering stack rather than just fine-tuning afterthoughts.

For security researchers, getting hands-on access to unreleased mitigation systems is a rare opportunity that is probably worth applying for before the August 16 deadline, even setting the cash bounty aside.