YFarmX Model Security CapabilityNewly scored models on the YFarmX Model Security Capability tracker: how each frontier model does on security work, under which harness and safeguard state, and what it will answer.https://yfarmx.com/en-GBMuse Spark: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/muse-spark/https://yfarmx.com/tools/model-security-capability/muse-spark/Muse Spark scores 65.4% end-to-end on Cybench on all 40 tasks, the highest full-task-set card figure on the leaderboard read 11 September 2026.Fri, 11 Sep 2026 00:00:00 GMTMetaapiCONFIRMEDClaude Opus 4.5: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/claude-opus-4-5/https://yfarmx.com/tools/model-security-capability/claude-opus-4-5/Claude Opus 4.5 scores 82% end-to-end on Cybench on the 39 of 40 tasks Anthropic ran, per the benchmark's own leaderboard read on 11 September 2026.Fri, 11 Sep 2026 00:00:00 GMTAnthropicapiCONFIRMEDClaude Opus 4.6: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/claude-opus-4-6/https://yfarmx.com/tools/model-security-capability/claude-opus-4-6/Claude Opus 4.6 scores 93% end-to-end on Cybench on the 37 of 40 tasks Anthropic ran, per the benchmark's own leaderboard read on 11 September 2026.Fri, 11 Sep 2026 00:00:00 GMTAnthropicapiCONFIRMEDClaude Opus 4.7: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/claude-opus-4-7/https://yfarmx.com/tools/model-security-capability/claude-opus-4-7/Claude Opus 4.7 scores 96% end-to-end on Cybench on the 35 of 40 tasks Anthropic ran, per the benchmark's own leaderboard read on 11 September 2026.Fri, 11 Sep 2026 00:00:00 GMTAnthropicapiCONFIRMEDClaude Mythos Preview: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/claude-mythos-preview/https://yfarmx.com/tools/model-security-capability/claude-mythos-preview/Claude Mythos Preview scores 100% end-to-end on Cybench on the 35 of 40 tasks Anthropic ran, per the benchmark's own leaderboard read on 11 September 2026. The task count is the honest way to state it: saturated on the subset the vendor ran.Fri, 11 Sep 2026 00:00:00 GMTAnthropicapiCONFIRMEDGemma 4 31B: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/gemma-4-31b/https://yfarmx.com/tools/model-security-capability/gemma-4-31b/The highest precision on the RealVuln board, 90.12%, at 23.50% recall and F3 25.4 on the Python subset, run on local hardware and billed at zero. A first-pass screen that reports few findings and is usually right about them.Fri, 11 Sep 2026 00:00:00 GMTGoogleopen-weightsCONFIRMEDQwen3.8-Max: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/qwen3-8-max/https://yfarmx.com/tools/model-security-capability/qwen3-8-max/A search-derived CVE campaign puts it at 26 of 32 recent CVEs, 81.25% pass@3, for $821.35 over three runs. Its smaller sibling Qwen3.8-27B ships under Apache-2.0 and had 150 refusal-removed rebuilds within a week of release.Fri, 11 Sep 2026 00:00:00 GMTAlibabaopen-weightsSINGLECyberKimi: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/cyberkimi/https://yfarmx.com/tools/model-security-capability/cyberkimi/Kimi K3 with the refusal layer removed and cyber post-training added. On the same V8 bug, harness and prompt as stock Kimi K3 it scored 8 of 16 capabilities unassisted and 10 of 16 with a methodology pack against the stock model's 4, the cleanest published control on how much willingness moves a security score.Fri, 11 Sep 2026 00:00:00 GMTAdverserial AIwaitlistCONFIRMEDKimi K3: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/kimi-k3/https://yfarmx.com/tools/model-security-capability/kimi-k3/The value pick on RealVuln: F3 59.4 on the Python subset at 73.02% precision for $21.33, $0.32 a repository. DeepSeek's table gives it 80.0 on CyberGym, and on ExploitBench against CVE-2024-6100 the stock model scored 4 of 16 capabilities. It is the model credited with a working Redis exploit, "the first llm that is capable and willing to write an exploit".Fri, 11 Sep 2026 00:00:00 GMTMoonshot AIopen-weightsCONFIRMEDDeepSeek V4 Pro: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/deepseek-v4-pro/https://yfarmx.com/tools/model-security-capability/deepseek-v4-pro/RealVuln scores it F3 38.0 across all 140 repositories at 60.91% precision over three runs; a search-derived CVE campaign puts it at 87.5% pass@3 at 65.6% precision on 32 recent CVEs.Fri, 11 Sep 2026 00:00:00 GMTDeepSeekopen-weightsCONFIRMEDDeepSeek V4 Flash: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/deepseek-v4-flash/https://yfarmx.com/tools/model-security-capability/deepseek-v4-flash/The cheapest measured pass on RealVuln: all 140 repositories at $0.0005 per 100 lines, F3 41.8 at 56.31% precision, $3.83 in total over three trials. A 120,000-line project costs about $0.62.Fri, 11 Sep 2026 00:00:00 GMTDeepSeekopen-weightsCONFIRMEDDeepSeek V4.1 Flash: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/deepseek-v4-1-flash/https://yfarmx.com/tools/model-security-capability/deepseek-v4-1-flash/Released 10 September 2026 under MIT. DeepSeek's own table gives 88.1 on CyberGym; RealVuln has it at F3 50.9 across all 140 repositories at 97.8 seconds a run, the fastest full-corpus wall clock on the board, billed at zero because it ran on local hardware.Fri, 11 Sep 2026 00:00:00 GMTDeepSeekopen-weightsCONFIRMEDGLM-5.3: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/glm-5-3/https://yfarmx.com/tools/model-security-capability/glm-5-3/Z.ai's own table gives 84.5 on CyberGym and 54.4 on ExploitBench; RealVuln has it at F3 56.9 across 136 of 140 repositories at $0.0148 per 100 lines, with four validation failures. It is the default model in the Strix pentesting agent's quickstart.Fri, 11 Sep 2026 00:00:00 GMTZ.aiopen-weightsCONFIRMEDGrok 4.6: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/grok-4-6/https://yfarmx.com/tools/model-security-capability/grok-4-6/On 28 August 2026 the Shannon agent ran Grok 4.6 through an authorised pentest of Photoview 2.4.0: 10 reported findings, 10 true positives, no false positives, for $35.07 in 5 hours 26 minutes, with the SARIF output opened by this desk. xAI is the one lab that publishes the text it injects in front of the model.Fri, 11 Sep 2026 00:00:00 GMTxAIapiCONFIRMEDGemini 3.5 Flash: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/gemini-3-5-flash/https://yfarmx.com/tools/model-security-capability/gemini-3-5-flash/The precision-shaped row on RealVuln: 89.77% precision at 33.26% recall, F3 35.5, on 64 of 140 repositories over three runs, with one timeout and one validation failure. It reports few findings and is usually right about them.Fri, 11 Sep 2026 00:00:00 GMTGoogleapiCONFIRMEDGemini 3.5 Flash Cyber: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/gemini-3-5-flash-cyber/https://yfarmx.com/tools/model-security-capability/gemini-3-5-flash-cyber/The first Flash Cyber model, credited by Google with 55 unique confirmed V8 issues on Big Sleep; its CyberGym figure is contested between 83.2% and 77.5% across Google's own materials.Fri, 11 Sep 2026 00:00:00 GMTGooglegatedCONFIRMEDGemini 3.8 Flash Cyber: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/gemini-3-8-flash-cyber/https://yfarmx.com/tools/model-security-capability/gemini-3-8-flash-cyber/The model inside CodeMender since 2 September 2026, reachable only through Fairwind. Google's own table gives 86.2% pass@1 on CyberGym, 47.2% on CWE-Bench at roughly $3.60 a rollout, and a 6.0% attack success rate on Gray Swan injection.Fri, 11 Sep 2026 00:00:00 GMTGooglegatedCONFIRMEDGPT-5.5: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/gpt-5-5/https://yfarmx.com/tools/model-security-capability/gpt-5-5/RealVuln scores GPT-5.5 at F3 56.7 on the Python subset at 72.62% precision over three runs, and Google's comparison table quotes GPT-5.5-Cyber at 85.6% pass@1 on CyberGym.Fri, 11 Sep 2026 00:00:00 GMTOpenAIapiCONFIRMEDGPT Daybreak Red: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/gpt-daybreak-red/https://yfarmx.com/tools/model-security-capability/gpt-daybreak-red/The only model in the field sold for "proof-of-concept exploit development, exploit-chain validation, penetration testing, and red teaming", at $12.50 and $75 per million tokens with a 400,000-token window. No public benchmark row exists for it.Fri, 11 Sep 2026 00:00:00 GMTOpenAIgatedSINGLEGPT Daybreak Blue: how it does on security work, and what it will answerhttps://yfarmx.com/tools/model-security-capability/gpt-daybreak-blue/https://yfarmx.com/tools/model-security-capability/gpt-daybreak-blue/The highest-scoring model row on RealVuln: F3 79.5 across all 140 repositories at 63.98% precision and 81.75% recall, $3.91 a repository, in 347.2 seconds a run. It is GPT-5.6 Sol's weights served to verified defenders under the Daybreak programme, and the 4.8-point gap to Sol on the same board is consistent with the permitted tier spending fewer turns declining.Fri, 11 Sep 2026 00:00:00 GMTOpenAIgatedCONFIRMED