{
  "id": 15742135,
  "title": "Exclusive: AI security company Gray Swan relaunches its red-teaming Arena contest for the era of AI agents",
  "url": "https://urgent.news/2026/10/20/exclusive-ai-security-company-gray-swan-relaunches-its-red-teaming",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-20T12:00:00.000Z",
  "source": {
    "name": "Fortune",
    "slug": "fortune",
    "url": "https://fortune.com/2026/10/20/exclusive-ai-security-company-gray-swan-red-team-arena-contest-to-jailbreak-ai-agents/"
  },
  "original_language": "en",
  "account": "AI security firm Gray Swan has revamped its red-teaming Arena platform to better suit the evolving landscape of AI agents. The Pittsburgh-based company, which assists frontier AI companies like Anthropic, OpenAI, and Google DeepMind in testing their models before release, introduced Arena 2.0 on Tuesday. Alongside this relaunch, Gray Swan is hosting its largest competition yet, the Indirect Prompt Injection (IPI) Black challenge, offering $300,000 in prizes from sponsors including Anthropic, Meta, Google DeepMind, and Reflection AI.\n\nThis new version of Arena will release distinct attack scenarios weekly until November 2. An indirect prompt injection is a security concern where an attacker embeds instructions within content an AI agent will encounter, aiming to manipulate the agent's behavior. Gray Swan's red-teaming platform includes simulated environments where these tests can be conducted safely without causing real-world damage.\n\nMatt Fredrikson, Gray Swan's CEO, explained that the platform consists of guarded versions of models for public use, while private arenas are available for frontier AI labs to test unguarded models with vetted participants who sign non-disclosure agreements. The company also builds realistic simulated environments, red-teams these environments internally and with third parties, and monitors agents' activities to identify potential zero-day vulnerabilities.\n\nThe original Arena, launched in September 2024, catered to chatbots, where contestants could jailbreak AI models relatively easily within minutes for open-source models and a day for models from companies like OpenAI and Anthropic. However, as AI labs have enhanced model robustness, the tests now require more technical expertise. As a result, the pool of capable and willing red teamers has shrunk, according to Fredrikson.\n\nDespite this, Gray Swan is committed to engaging its community of citizen red teamers. They offer beginner-friendly challenges, training certificates, and a pathway for top performers to join their red-teaming community for real-world testing engagements. Fredrikson emphasized that understanding AI system safety and security is crucial, and it shouldn't be confined to a small circle of specialists.",
  "summary": "Gray Swan's new Arena includes $300,000 in prize money for developing tactics to jailbreak models from leading AI companies",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}