Google DeepMind Launches Gemini 3.5 Flash Cyber for Vulnerability
SkimNews Take
Bundling autonomous vulnerability discovery with same-agent patching transforms cyber AI from an advisor into an active producer of code changes, making the model's restricted access itself a defensive advantage.
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Google DeepMind announced Gemini 3.5 Flash Cyber, a model built atop 3.5 Flash designed to discover, validate, and patch vulnerabilities, exclusively available to governments and trusted partners via CodeMender in a limited-access pilot
- 3.5 Flash Cyber is positioned as a cost-efficient lightweight alternative to large cybersecurity models, letting CodeMender scan more code paths at high speed and low cost, with plans to add red-teaming and end-to-end enterprise defense
- DeepMind's internal evaluations found the model outperformed Gemini 3.5 Flash, Gemini 3.6 Flash, and Anthropic Claude Opus 4.6, uncovering 55 unique confirmed issues on the V8 JavaScript Engine versus 47 and 36 respectively — including 10 issues no other model caught
- The model produced a 100% reliable remote-code execution exploit that bypassed standard mitigations like Address Space Layout Randomization (ASLR) and Write XOR Execute (W^X), and was used internally to find remote code execution vulnerabilities in public APIs and a memory-corruption flaw in a production service
- The launch shipped alongside Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, optimized for coding, knowledge work, multimodal performance, and low-latency tasks respectively
- Google is separately bringing CodeMender's foundational capabilities to enterprise customers through the generally available Gemini Enterprise Agent Platform
Why it matters: DeepMind is gating its best vulnerability-finding model behind a governments-and-trusted-partners pilot while shipping a commercial twin through Gemini Enterprise — a dual track that gives defenders an edge on zero-days but concentrates cutting-edge offensive-grade capability in vetted hands, with Claude Opus 4.6 as the named benchmark competitor that found 19 fewer V8 issues.
