Atlas

Benchmarks

← All benchmarks

dfbench Detection Recall

Cybersecurity · 2026-08-19

depthfirst's held-out defensive security benchmark, built from real vulnerabilities in open-source projects. An agent audits a repository scope in depthfirst's internal harness with network access disabled, and detection recall is the share of the scope's known vulnerabilities it finds. The August 19, 2026 post describes 253 examples and 910 vulnerabilities, and the page says the dataset will keep evolving. depthfirst ran these evaluations under providers' trusted access programs without saying which rows did, so scores aren't established as the stock models' and stay diagnostic.

Benchmark source: Depthfirst

Top models (higher is better)

ModelScoreSource
GPT-5.6 Sol65.7dfbench: GPT 5.6 Sol xhigh · depthfirst
GPT-6 Sol59.3dfbench: GPT 6 Sol xhigh · depthfirst
Grok 4.759.0dfbench: Grok 4.7 xhigh · depthfirst
DeepSeek V4.1 Flash57.3dfbench: DeepSeek v4.1 Flash high · depthfirst
Grok 4.557.2dfbench: Grok 4.5 high · depthfirst
MiMo-V2.6-Pro55.0dfbench: MiMo-v2.6 Pro · depthfirst
Claude Opus 5.554.9dfbench: Opus 5.5 high · depthfirst
MiMo-V2.6-Flash53.5dfbench: MiMo-v2.6 Flash · depthfirst
GPT-5.6 Luna52.4dfbench: GPT 5.6 Luna xhigh · depthfirst
Kimi K348.0dfbench: Kimi K3 max · depthfirst
Claude Opus 547.8dfbench: Opus 5 med · depthfirst
Muse Spark 1.347.4dfbench: Muse Spark 1.3 xhigh · depthfirst
GPT-6 Luna45.7dfbench: GPT 6 Luna xhigh · depthfirst
GLM-5.240.7dfbench: GLM 5.2 xhigh · depthfirst
Gemini 3.6 Flash20.9dfbench: Gemini 3.6 Flash high · depthfirst
Loading Atlas data…