Google gates its sharpest cyber model, joins Anthropic and OpenAI playbook

Google is not selling its strongest cybersecurity model on the open market. On 2 September, alongside making Gemini 3.8 Flash generally available through its API, Google launched Gemini 3.8 Flash Cyber, a security-tuned variant that only ships through a new vetted-access scheme called the Fairwind Program. That gating decision, not the model itself, is the commercial story.

Gemini 3.8 Flash Cyber is a post-trained successor to 3.5 Flash Cyber, built on the same underlying intelligence as the standard release but with fewer guardrails on advanced security tasks, letting it perform work the general-purpose model would block. It searches code for vulnerabilities, verifies the findings, generates fixes and validates patches before deployment. Google pairs the model with its CodeMender agent harness, which it says helps defenders "find, verify, and fix vulnerabilities at agentic scale."

Access runs through application only, with no self-serve tier. Google is prioritising government agencies, Google Cloud customers and security partners, with particular weight given to critical-infrastructure operators and maintainers of widely used software. More than 650 organisations are already in the programme, and members must restrict access to security staff and use protections such as multi-factor authentication. Cloud customers who don't qualify can still use CodeMender with public models, just without the Cyber variant's loosened restrictions.

On performance, Google cites internal tests showing over 70% vulnerability discovery across 20 languages and 2.6 times more correct patches than competing models. On a Collinear patching benchmark, the model lands on the cost-performance frontier: 47.2% pass@1 against 47.8% for a leading frontier model, at much lower cost, according to Google.

The move follows Anthropic, which gated Claude Mythos 5 behind its Project Glasswing programme in April, and OpenAI, which restricted GPT-5.6-Cyber through its Daybreak programme, according to Security Boulevard.

Three major labs now agree that the most capable offensive-grade cyber models are too dangerous to sell freely, and all three have converged on the same fix: an application-gated programme rather than an API toggle. That consensus effectively sets an industry norm competitors will have to match to win critical-infrastructure and government business, since buyers in that segment will expect vetting as standard rather than a differentiator.

For Google, the scale of Fairwind's rollout, more than 650 organisations before public marketing, suggests it moved with real enterprise and government demand already lined up rather than testing appetite. The cost-performance figures against a leading frontier model matter more than the raw discovery rate: undercutting a rival on price while matching it on capability is the lever that pulls infrastructure operators and Cloud customers away from competing vendors, provided the vetting process doesn't slow adoption enough to blunt that advantage.