Frontier Log

Benchmark · Active

ExploitGym

Created by Berkeley RDI, 11 May 2026 · Higher is better · Website

Settings

Task setScoringMitigationsReported by
ExploitGym 1.0 (869 tasks)ScoreMitigations offOpenAIGoogle
ExploitGym 1.0 (869 tasks)ScoreMitigations onOpenAI
ExploitGym 1.0 (869 tasks)ScoreNot specifiedOpenAIMetaDeepSeekMoonshot AI
ExploitGym 1.0 (869 tasks)Successful exploitsMitigations offAnthropicMetaZ.ai
ExploitGym 1.0 (869 tasks)Successful exploitsNot specifiedMeta

Bold: Benchmark default based on Paper and Leaderboard · Download image

Results

Show
0200400600869Jul 2026Oct 2026341Mythos 5.2130GLM-5.382Haiku 5.5231Sonnet 5.5300Opus 5.5264Mythos 5.1191Opus 551Sonnet 50200400600869Jul 2026Oct 2026341Mythos 5.2130GLM-5.382Haiku 5.5231Sonnet 5.5300Opus 5.5
  • Anthropic
  • Other labs
  • Reported in a later post or card
ModelHarness and effortScoreSource
Claude Mythos 5.29 Nov 2026Harness not specifiedEffort not specified341 of 869Claude Fable 5.2 and Claude Mythos 5.2 System CardChecked A person checked this number against its source on 20 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.6h wall-clock budget · agent judge confirms the intended vulnerabilityOur capture
Claude Mythos 5.29 Nov 2026Harness not specifiedEffort not specified331 of 869Claude Fable 5.2 and Claude Mythos 5.2 System CardChecked A person checked this number against its source on 20 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.2h wall-clock budget · agent judge confirms the intended vulnerability · another reportOur capture
Claude Opus 5.522 Sep 2026Harness not specifiedEffort not specified300 of 869Claude Opus 5.5 System CardCorroborated The same number is printed in 3 other documents: 1 2 36h wall-clock budget · agent judge confirms the intended vulnerabilityOur capture
Claude Opus 5.522 Sep 2026Harness not specifiedEffort not specified289 of 869Claude Opus 5.5 System CardCorroborated The same number is printed in 1 other document: 12h wall-clock budget · agent judge confirms the intended vulnerability · another reportOur capture
Claude Mythos 5.11 Sep 2026Harness not specifiedEffort not specified264 of 869Claude Fable 5.1 and Claude Mythos 5.1 System CardCorroborated The same number is printed in 2 other documents: 1 26h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · Setting from Berkeley RDI’s leaderboardOur capture
Claude Mythos 5.11 Sep 2026Harness not specifiedEffort not specified203 of 869Claude Fable 5.1 and Claude Mythos 5.1 System CardCorroborated The same number is printed in 2 other documents: 1 22h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · another report · Setting from Berkeley RDI’s leaderboardOur capture
Claude Mythos 59 Jun 2026Harness not specifiedEffort not specified247 of 869Claude Opus 5 System CardCorroborated The same number is printed in 2 other documents: 1 26h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · reported later · Setting from Berkeley RDI’s leaderboardOur capture
Claude Mythos 59 Jun 2026Harness not specifiedEffort not specified181 of 869Claude Opus 5 System CardCorroborated The same number is printed in 2 other documents: 1 22h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · reported later · another report · Setting from Berkeley RDI’s leaderboardOur capture
Claude Sonnet 5.528 Sep 2026Harness not specifiedEffort not specified231 of 869Claude Sonnet 5.5 System CardCorroborated The same number is printed in 1 other document: 16h wall-clock budget · agent judge confirms the intended vulnerabilityOur capture
Claude Sonnet 5.528 Sep 2026Harness not specifiedEffort not specified205 of 869Claude Sonnet 5.5 System CardCorroborated The same number is printed in 1 other document: 12h wall-clock budget · agent judge confirms the intended vulnerability · another reportOur capture
Claude Opus 524 Jul 2026Harness not specifiedEffort not specified191 of 869Claude Opus 5 System CardCorroborated The same number is printed in 3 other documents: 1 2 36h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · Setting from Berkeley RDI’s leaderboardOur capture
Claude Opus 524 Jul 2026Harness not specifiedEffort not specified171 of 869Claude Opus 5 System CardCorroborated The same number is printed in 3 other documents: 1 2 32h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · another report · Setting from Berkeley RDI’s leaderboardOur capture
GLM-5.314 Aug 2026Harness not specifiedmax reasoning effort130 of 869GLM-5.3: Frontier Coding with Emergent Cyber CapabilitiesCorroborated The same number is printed in 1 other document: 1single-run pass@1 · no web tools, domain whitelist · 6h budget, API inference time rescaled by the model's tokens per second, plus non-API overhead · Setting from Berkeley RDI’s leaderboardOur capture
GLM-5.314 Aug 2026Harness not specifiedmax reasoning effort105 of 869GLM-5.3: Frontier Coding with Emergent Cyber CapabilitiesCorroborated The same number is printed in 1 other document: 1single-run pass@1 · no web tools, domain whitelist · 2h budget, API inference time rescaled by the model's tokens per second, plus non-API overhead · another report · Setting from Berkeley RDI’s leaderboardOur capture
Claude Opus 4.828 May 2026Harness not specifiedEffort not specified120 of 869Claude Opus 5 System CardCorroborated The same number is printed in 2 other documents: 1 26h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · reported later · Setting from Berkeley RDI’s leaderboardOur capture
Claude Opus 4.828 May 2026Harness not specifiedEffort not specified80 of 869Claude Opus 5 System CardCorroborated The same number is printed in 2 other documents: 1 22h wall-clock budget · agent judge confirms the intended vulnerability, the judge is the model being evaluated · reported later · another report · Setting from Berkeley RDI’s leaderboardOur capture
Claude Haiku 5.57 Oct 2026Harness not specifiedEffort not specified82 of 869Claude Haiku 5.5 System CardCorroborated The same number is printed in 1 other document: 16h wall-clock budget · agent judge confirms the intended vulnerabilityOur capture
Claude Haiku 5.57 Oct 2026Harness not specifiedEffort not specified79 of 869Claude Haiku 5.5 System CardCorroborated The same number is printed in 1 other document: 12h wall-clock budget · agent judge confirms the intended vulnerability · another reportOur capture
Claude Sonnet 530 Jun 2026Harness not specifiedEffort not specified51 of 869Claude Opus 5.5 System CardCorroborated The same number is printed in 3 other documents: 1 2 36h wall-clock budget · agent judge confirms the intended vulnerability · reported laterOur capture
Claude Sonnet 530 Jun 2026Harness not specifiedEffort not specified25 of 869Claude Opus 5.5 System CardCorroborated The same number is printed in 3 other documents: 1 2 32h wall-clock budget · agent judge confirms the intended vulnerability · reported later · another reportOur capture
GLM-5.216 Jun 2026Harness not specifiedEffort not specified39 of 869GLM-5.3: Frontier Coding with Emergent Cyber CapabilitiesCorroborated The same number is printed in 1 other document: 16h budget · reported later · Setting from Berkeley RDI’s leaderboardOur capture
GLM-5.216 Jun 2026Harness not specifiedEffort not specified29 of 869GLM-5.3: Frontier Coding with Emergent Cyber CapabilitiesCorroborated The same number is printed in 1 other document: 12h budget · reported later · another report · Setting from Berkeley RDI’s leaderboardOur capture
Muse Spark 1.19 Jul 2026Harness not specifiedEffort not specified7 of 869Muse Spark 1.1 Evaluation ReportChecked A person checked this number against its source on 20 Nov 2026. This confirms the number matches what the lab published. It doesn't confirm that the score was reproduced.4h timeout · Setting from Berkeley RDI’s leaderboard

All benchmarks