Tools Model Security Capability gemini-3-5-flash-cyber
Model record
Gemini 3.5 Flash Cyber: how it does on security work, and what it will answer
- Lab
- Released
- 2026-07-21
- Access
- gated
- Weights
- closed
- Safeguards
- Limited-access pilot through CodeMender for governments and trusted partners
- Score rows
- 2
- Confidence
- CONFIRMED
What happened
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.
Big Sleep on V8, unique confirmed issues: 55 (vendor-card; Google Big Sleep harness; Chromium V8; safeguards: cyber-tuned model inside Google's own agent; run 2026-07-21; CONFIRMED). CyberGym, pass@1: 83.2 or 77.5% (vendor-card; Google launch materials, 21 July 2026; CyberGym; safeguards: cyber-tuned model; run 2026-07-21; CONTESTED). two figures in Google's own materials. Refusal gate (SINGLE, observed 2026-09-11): Limited-access pilot; Google mitigates model abuse by disabling malicious accounts.
Sources
- Big Sleep on V8 sourceyfarmx.com/ai/llms/gemini-3-5-family/
- CyberGym sourceyfarmx.com/ai/llms/gemini-3-5-family/
- Refusal policy or licencecloud.google.com/blog/topics/threat-intelligence/ai-vulnerab…
On YFarmX
- Reference pageyfarmx.com/ai/llms/gemini-3-5-family/
One record from the Model Security Capability, maintained by the Security Desk. Data: CSV · JSON ·RSS · CC BY 4.0 with attribution to YFarmX.Tracker updated · 19 September 2026