(Disclosure: I work at Gray Swan, where I run the Frontier Lab Safety team and the Arena)

We just launched our largest-ever AI red teaming challenge, at the request of Anthropic, Meta, Google DeepMind, Reflection, and our other lab sponsors, who want to see just how strong their auto mode defenses are and to what degree prompt injections on coding agents are solved problems.

It's been live for two days, but we need more breaks! --which means the fruit hang low; right now it's only 4 breaks out of 84 targets to get on the prized leaderboard for white box, and even 1 break out of 52 on black box would currently get you fifth place ($2500).

Experience with AI red teaming is helpful, but for this, general bug bounty experience is probably more important, so imagine some of y'all would be quite good at these kinds of exploits.

Check it out:…