Solutions  ·  2026-10-23

Gray Swan relaunches Arena 2.0 — public red-teaming platform rebuilt for AI agents

SolutionsMedium impactGlobal
On Oct 20 Gray Swan launched Arena 2.0, a redesigned public red-teaming platform, and opened its largest competition yet: 'Indirect Prompt Injection (IPI) Black', a $300K challenge sponsored by Anthropic, Meta, Google DeepMind and Reflection AI, with new attack scenarios weekly through Nov 2. The platform builds realistic simulated environments with sandboxes, requires a human to drive each agent, and adds a graduated learning path, training certificates, and a route into Gray Swan's paid red-team community.
IPI via poisoned web content/forums is the dominant agent risk, and this brings structured, sponsored agent-focused red-teaming into the public domain with safe sandboxed execution. Red-teaming is moving from static model jailbreaks to multi-turn agent workflow escapades — a leading-edge signal for agent security testing.
Security researchers, red teams, and enterprises building web-using agents should watch/participate to understand IPI attack patterns and agent sandbox escape behavior; useful for evaluating agent red-teaming platforms before purchase.
Fortune exclusive on Gray Swan Arena 2.0
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →