Google has built a cybersecurity model capable of finding vulnerabilities, checking whether they can be exploited, and generating patches. Most organizations will not be allowed to use it.
Google DeepMind announced Gemini 3.5 Flash Cyber on July 21, 2026, with initial access reserved for governments and trusted partners through CodeMender. Google has not published pricing, selection criteria, a standalone API, or a general-release date. A separate CodeMender preview is available to selected enterprise customers, but only for testing and evaluation — not commercial or production use.
How Flash Cyber searches for flaws
Built on Gemini 3.5 Flash, the specialized model is tuned for vulnerability discovery, validation, and patching. Google said in its model announcement that Flash Cyber is already running through CodeMender across internal Chrome, Android, Google Cloud, Ads, and YouTube codebases.
CodeMender is also available to a limited set of customers using generally available Gemini models. Its current documentation lists Gemini 3.5 Flash, Gemini 3.1 Pro Preview, and Gemini 3 Flash Preview, but not Flash Cyber.
For the CyberGym benchmark, Google configured CodeMender to call Flash Cyber as many as five times before producing one report. The company says repeated calls to a lightweight model could support more frequent scans than a single invocation of a larger model.
In a V8 JavaScript engine test, Flash Cyber found 55 unique confirmed issues. Standard Gemini 3.5 Flash found 47, while Claude Opus 4.6 found 36. Google said 10 issues were missed by both comparison models.
Other AI-assisted vulnerability pipelines have shown how automation can expand API mapping and access-control testing, although reported results still need independent verification.
Google’s figures also have limits. Some CyberGym competitor scores were self-reported by their providers, and Google has not published false-positive rates, patch-acceptance rates, regression data, or independent production testing.
The restricted rollout reflects dual-use risk. Technology that helps defenders find software flaws could also help attackers, as shown by an AI-assisted zero-day exploit Google documented in May 2026.
The controls CodeMender still needs
Mandiant’s M-Trends 2026 report estimates a mean time to exploit of negative seven days, meaning some vulnerabilities were exploited before patches became available. Google has not shown that CodeMender shortens that interval in external production environments.
Before testing the preview, organizations should verify pricing and access terms, external performance evidence, code-handling rules, and preview restrictions. CodeMender’s hosted reasoning and local execution model sends selected code snippets, proposed patches, and command results to Google’s cloud systems while builds and exploit checks run locally.
Google says full repositories are not uploaded and customer code is not used to train the underlying models. Session data can be retained for up to seven days, with earlier deletion available through an API call.
CodeMender should run in an isolated environment with human review, regression testing, and rollback controls for generated patches. Repository and tool permissions also require strict limits because AI agents can create security gaps when access exceeds their assigned task. Until Google publishes production evidence and broader access terms, enterprises should evaluate the workflow and its controls — not budget around the specialized model.
Read more: Code-scanning agents also need protection from the data they consume because a trusted AI assistant can become an attack path when repositories contain hidden or malicious instructions.