XBOW's early access testing shows GPT-5.5 matches Anthropic's restricted Mythos model on penetration testing benchmarks — and OpenAI is releasing it publicly.
XBOW, an AI-powered offensive security firm, conducted early access testing of OpenAI's GPT-5.5 inside their agent-based penetration testing workflows. The model delivered a measurable step-change in vulnerability detection and exploitation tasks — performance they describe as comparable to Anthropic's Mythos, which remains restricted. Unlike Mythos, GPT-5.5 is being released broadly. XBOW reports GPT-5.5 is now leading across several of their internal benchmarks, describing it as 'generally more powerful' than prior model generations.
GPT-5.5 isn't just incrementally better — XBOW's agent-in-the-loop benchmarks show it outperforming previous frontier models on multi-step offensive security tasks including login flows, vuln discovery, and report generation. This isn't a chat benchmark; it's performance inside a real agentic stack. For developers building security tooling or agent pipelines, this is a meaningful capability jump available via the standard OpenAI API.
If you're running any automated security scanning or agent workflow on OpenAI's API, swap your current model config to GPT-5.5 and run your existing test suite — XBOW's results suggest you'll see measurable accuracy gains on multi-step reasoning tasks without re-architecting anything.
Open your terminal and ensure your OPENAI_API_KEY is set
Tags