Get Paid to Manipulate AI 😈

Ever heard of ‘AI red-teaming’?

It's the practice of intentionally trying to break an AI system, get it to say something harmful, leak information it shouldn't, or take an action it's not supposed to take so the people building it can fix the weakness before the public ever encounters it.

Companies building frontier AI models (OpenAI, Anthropic, Google, Amazon, Meta) pay real people to try to break their systems before the public ever sees them. Gray Swan Arena is the platform where a lot of that work happens, and it's open to anyone 18+.

Gray Swan runs competitive challenges where participants try to manipulate AI models into doing things they shouldn't. This includes things like jailbreaking, prompt injection, and agent manipulation. You don’t need a degree or coding background to do this. Most successful attempts at breaking these models come from creative prompting and clear writing, not programming skill.

Major competitions offer prize pools from $20,000 to $170,000 or more. In one recent challenge focused on AI agents, participants collectively ran 1.8 million attack attempts, landed 62,000 successful breaches, and split $171,800 in payouts. Weekly challenges pay smaller amounts but run constantly, so newer participants can earn while they learn.

Companies use this data to decide whether a model is ready to ship and what needs fixing first. Data is gold, including data on weaknesses and vulnerabilities. That $171,800 payout came from a challenge co-sponsored by UK AISI, OpenAI, Anthropic, and Google DeepMind. Out of thousands of participants, 161 walked away with cash prizes.

Participating and being a top performer can also lead to a job. Top performers get pulled into Gray Swan's private red-teaming network, which pays for direct engagements with AI labs. Gray Swan has hired more than 14 Arena winners onto its own team. Independent researcher Ophira Horwitz built her reputation this way, first by exposing a vulnerability in Anthropic's Claude Sonnet-3.5, then by becoming one of only two competitors to crack a Gray Swan model called Cygnet, using playful, upbeat prompts to get past its defenses. She and the other competitor who cracked it both walked away with cash bounties and were later hired as Gray Swan consultants.

The timeline on breaking into a career varies for all participants. One participant told me it took about 12 months of steady work before getting invited into the private network, but others get in faster depending on results.

🔗 Sign up at https://app.grayswan.ai/arena.

Proving Ground runs new challenges every Wednesday across categories like Chat, Image, Agent, and Indirect prompt injection, and it's built for people who are still building up their skills. Every successful break there adds a verified credential to your public profile, which is what labs and companies actually look at when scouting for paid gigs. The Discord runs alongside it where new competitions and payout announcements post before anywhere else.

If you’ve been looking for a way into AI security, this is one of the most direct and accessible paths available right now.