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
0102030405060Apr 2026Jul 2026Oct 202648.5%GPT-6-Cyber20.7%Kimi K3.1 Pro42.4%GPT-6 Astra33.7%GPT-5.6 Sol21.2%GPT-5.5-Cyber0102030405060Apr 2026Jul 2026Oct 202648.5%GPT-6-Cyber20.7%Kimi K3.1 Pro42.4%GPT-6 Astra33.7%GPT-5.6 Sol21.2%GPT-5.5-Cyber
  • OpenAI
  • Other labs
  • Reported in a later post or card
ModelHarness and effortScoreSource
GPT-6-Cyber5 Nov 2026Harness not specifiedEffort not specified48.5%GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agentsChecked 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.
GPT-6 Astra3 Sep 2026Harness not specifiedEffort not specified42.4%GPT-6 Astra: A new generation of intelligenceCorroborated The same number is printed in 4 other documents. The first 3 are linked: 1 2 3no 6h time limit, output-token cap per the system card · "v1" implementation with runtime package installation disabled and the environment offline, per the system card · Daybreak Blue access, per the GPT-6 Cyber post · without production safeguardsOur capture
GPT-6 Astra3 Sep 2026Harness not specifiedEffort not specified0.0%GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agentsChecked 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.strict supplied-vulnerability pass@1 · safeguards enabled · reported later · another reportOur capture
GPT-6.1 Sol29 Sep 2026Harness not specifiedEffort not specified35.1%GPT-6.1 Sol System Card AddendumCorroborated The same number is printed in 1 other document: 1per attempt
GPT-5.6 Sol9 Jul 2026Harness not specifiedEffort not specified33.7%GPT-5.6: Frontier intelligence that scales with your ambitionCorroborated The same number is printed in 1 other document: 16h time limit, timed on OpenAI's faster alpha API and rescaled to the public API · reduced safeguardsOur capture
GPT-5.6 Sol9 Jul 2026Harness not specifiedEffort not specified24.9%GPT-5.6: Frontier intelligence that scales with your ambitionChecked 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 cap, timed on OpenAI's faster alpha API and rescaled to the public API · reduced safeguards · another report
GPT-5.6 Sol9 Jul 2026Harness not specifiedEffort not specified30.3%GPT-6 Astra: A new generation of intelligenceCorroborated The same number is printed in 3 other documents: 1 2 3no 6h time limit · Daybreak Blue access, per the GPT-6 Cyber post · system-level guardrails removed, with safeguards tailored to authorized defensive security work, per the GPT-5.6-Cyber post · reported later · another reportOur capture
GPT-5.6 Sol9 Jul 2026Harness not specifiedhighest publicly available reasoning level30.8%Expanding Daybreak as the Cyber Defense Window NarrowsChecked 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.Daybreak Blue access · system-level guardrails removed, with safeguards tailored to authorized defensive security work · OpenAI's internal implementation in security-hardened, isolated environments · reported later · another report
GPT-5.6 Sol9 Jul 2026Harness not specifiedhighest publicly available reasoning level0.0%Expanding Daybreak as the Cyber Defense Window NarrowsChecked 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.safeguards enabled · OpenAI's internal implementation in security-hardened, isolated environments · reported later · another report
GPT-5.6-Cyber10 Aug 2026Harness not specifiedhighest publicly available reasoning level32.1%Expanding Daybreak as the Cyber Defense Window NarrowsCorroborated The same number is printed in 1 other document: 1Daybreak Red access · OpenAI's internal implementation in security-hardened, isolated environments
GPT-5.6 Terra9 Jul 2026Harness not specifiedEffort not specified23.2%GPT-5.6: Frontier intelligence that scales with your ambitionChecked 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.timed on OpenAI's faster alpha API and rescaled to the public API · reduced safeguardsOur capture
GPT-6 Sol22 Sep 2026Harness not specifiedEffort not specified22.1%GPT-6.1 Sol System Card AddendumCorroborated The same number is printed in 1 other document: 1reported later
GPT-5.5-Cyber7 May 2026Harness not specifiedhighest publicly available reasoning level21.2%Expanding Daybreak as the Cyber Defense Window NarrowsChecked 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.Daybreak Red access · OpenAI's internal implementation in security-hardened, isolated environments · reported later
Kimi K3.1 Pro24 Oct 2026KimiCodemax effort20.7%Kimi K3.1 Pro Tech Blog: Advancing Open IntelligenceChecked 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 per taskOur capture
DeepSeek-V4.1-Flash10 Sep 2026Harness not specifiedEffort not specified15.3%DeepSeek-V4.1-Flash: Smarter, Faster, More EfficientChecked 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.Source image
GPT-5.523 Apr 2026Harness not specifiedEffort not specified15.1%GPT-5.6: Frontier intelligence that scales with your ambitionChecked 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 cap, timed on OpenAI's faster alpha API and rescaled to the public API · reduced safeguards · reported laterOur capture
GPT-5.6 Luna9 Jul 2026Harness not specifiedEffort not specified12.4%GPT-5.6: Frontier intelligence that scales with your ambitionChecked 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.timed on OpenAI's faster alpha API and rescaled to the public API · reduced safeguardsOur capture
GPT-6 Luna22 Sep 2026Harness not specifiedEffort not specified11.6%GPT-6.1 Sol System Card AddendumCorroborated The same number is printed in 1 other document: 1reported later
DeepSeek-V4-Pro-081313 Aug 2026Harness not specifiedEffort not specified5.4%DeepSeek-V4.1-Flash: Smarter, Faster, More EfficientChecked 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.reported laterSource image
Muse Spark 1.19 Jul 2026Harness not specifiedhigh reasoning effort0.8%Muse 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.pass@1 · Muse Spark 1.1 Helpful, no system prompt

All benchmarks