
Kudos to lead researcher @devleibowitz 🎉 and the whole Novee AI Research team led by @DanPadnos
✓Building autonomous offensive security agents. Trained to continuously find, prove, and fix high-impact vulnerabilities before attackers can act.

Kudos to lead researcher @devleibowitz 🎉 and the whole Novee AI Research team led by @DanPadnos

Dangerbench: we ran our pentesting agent on live apps across 5 frontier models, with and without our layered guardrails. Safety violations: 85 → 0 Verified findings: 169 → 175 The safety-vs-capability tradeoff didn’t show up. Full method and chart: novee.security/blog/safety-i…

1/ Is it risky to pentest your app with AI? We put 5 frontier models in a pentesting harness and pointed them at live webapps. Each ran twice: unguarded, then with our guardrail stack. Safety violations: 85 → 0 Verified findings: 169 → 176 No safety-vs-capability tradeoff!

We put PWNBench on @FireworksAI_HQ Specialized Intelligence Index, alongside Harvey (legal) and Doximity (healthcare), so model builders and security teams can compare models on live offensive work. See the results: fireworks.ai/specialized-int… Full method: novee.security/blog/ai-pente…

PWNBench is built on live web applications with verified, ground-truth vulnerabilities. Agents have to explore noisy targets, decide where to focus, and prove real risk. Then we score recall, precision, and cost. It’s not a CTF puzzle, it’s not a bug bounty competition.

Most AI benchmarks score bounded, static tasks. That doesn’t match up with offensive security ops. We should be asking, can the model break into a live system it’s never seen and prove what’s exploitable? We built a benchmark that measures it the right way: PWNBench.