evals.report
BenchmarksLabsCompareRun guidesIn the wild
BenchmarksAgents

Gray Swan Arena (Agent Red-Teaming / Indirect Prompt Injection)

A large-scale public red-teaming competition run on the Gray Swan Arena platform that measures how often adversarial attackers can break frontier AI agents (via jailbreaks and indirect prompt injection across tool-use, coding, and computer-use settings), reported as an attack success rate where lower is better.

AgentsAttack Success Rate (ASR)Lower is better
ScoresAboutRun this benchmark

What is Gray Swan Arena (Agent Red-Teaming / Indirect Prompt Injection)?

A large-scale public red-teaming competition run on the Gray Swan Arena platform that measures how often adversarial attackers can break frontier AI agents (via jailbreaks and indirect prompt injection across tool-use, coding, and computer-use settings), reported as an attack success rate where lower is better. evals.report tracks reported Gray Swan Arena (Agent Red-Teaming / Indirect Prompt Injection) scores with the model, source, status, date, and run caveats attached — official leaderboard scores, vendor-reported launches, and clearly labeled community runs.

Top reported Gray Swan Arena (Agent Red-Teaming / Indirect Prompt Injection) score: Claude Opus 4.5 — 0.5% (Attack Success Rate (ASR)).

ModelLabScore↓Source modelStatusDate
Gemini 2.5 ProGoogle DeepMind8.5%—VerifiedMar 25, 2025Details
Llama 3.1 405BOpenMeta5.89%—VerifiedJul 23, 2024Details
DeepSeek V3.1OpenDeepSeek5.4%—VerifiedAug 21, 2025Details
Kimi K2 ThinkingOpenMoonshot AI4.8%—VerifiedNov 6, 2025Details
Grok 4xAI2.9%—VerifiedJul 9, 2025Details
GPT-5.1OpenAI2.5%—VerifiedNov 12, 2025Details
o3OpenAI2.50%—VerifiedApr 16, 2025Details
GPT-4oOpenAI2.41%—VerifiedMay 13, 2024Details
GPT-5OpenAI2.0%—VerifiedAug 7, 2025Details
Claude 3.5 SonnetAnthropic1.85%—VerifiedJun 20, 2024Details
Claude 3.7 SonnetAnthropic1.61%—VerifiedFeb 24, 2025Details
Claude Haiku 4.5Anthropic1.3%—VerifiedOct 15, 2025Details
Claude Sonnet 4.5Anthropic1.0%—VerifiedSep 29, 2025Details
Claude Opus 4.5Anthropic0.5%—VerifiedNov 24, 2025Details

Each row reports the model’s Attack Success Rate (ASR) on Gray Swan Arena (Agent Red-Teaming / Indirect Prompt Injection). Click a row for the full run context.