Google launches Gemini 3.5 Flash Cyber for vulnerability hunting — but keeps it on a tight leash

What happened

Google announced three new Gemini Flash models on July 21, 2026, headlined by Gemini 3.5 Flash Cyber, a security-tuned model designed to find, validate, and patch software vulnerabilities. The model is paired with Google's CodeMender code-security agent, which entered preview the same day.

According to a dedicated DeepMind blog post, Flash Cyber is built on top of the base 3.5 Flash model and fine-tuned specifically for vulnerability discovery and patching, making it more effective at those tasks than Gemini's mainline Flash models. Google positions it as a "cost-efficient and highly capable alternative" to larger, more expensive AI security systems such as Anthropic's Mythos.

However, Google is deliberately limiting who can use it. Citing the "dual-use nature" of the technology, the company is making Flash Cyber available exclusively to governments and trusted partners via CodeMender through a limited-access pilot program. DeepMind says the program will be "expanding over time," but gave no timeline for broader availability.

The launch also included two sibling models: Gemini 3.6 Flash, a workhorse model that Google says uses 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index and is priced at $1.50 per million input tokens and $7.50 per million output tokens; and Gemini 3.5 Flash-Lite, described as the fastest and most cost-effective 3.5-class model, delivering 350 output tokens per second.

Notably absent from the announcement was Gemini 3.5 Pro, Google's flagship model. Originally slated for a June launch per CEO Sundar Pichai's I/O keynote in May, the Pro model remains in testing with partners. Google said it will ship "as soon as it is ready" but offered no updated timing. Meanwhile, The Decoder reports that pretraining for Gemini 4 is already underway, which Google calls its "most ambitious training run" yet.

Why it matters

The cybersecurity AI market is heating up. Anthropic's Mythos models have established a foothold in AI-assisted vulnerability discovery, and Google's entry with a Flash-tier model signals a bet that specialized, efficient models — not just the largest frontier systems — can compete in security workloads.

The restricted access is the key tension. By limiting Flash Cyber to governments and trusted partners, Google is acknowledging that a model skilled at finding vulnerabilities is also skilled at exploiting them. This is a more cautious deployment posture than typical API releases, and it means most developers and security teams cannot simply call the model today. The pilot's planned expansion offers a path forward, but without a timeline, the broader market is left waiting.

The continued absence of Gemini 3.5 Pro adds another layer of pressure. With Anthropic and OpenAI pushing ahead with frontier models, Google's inability to ship its flagship — now delayed by at least a month — raises questions about whether DeepMind can keep pace. The Gemini 4 pretraining disclosure reads as a forward-looking signal, but it also underscores that the current generation's top model is still missing.

What to watch

• Pilot expansion timeline: Google says the Flash Cyber pilot will expand "over time" but has not specified when or to whom. Watch for updates on eligibility criteria.
• Gemini 3.5 Pro launch: Every Flash release without Pro deepens the narrative of a delayed flagship. Google's next announcement cycle will be scrutinized for a ship date.
• Independent benchmarks: Google claims Flash Cyber delivers "competitive performance at the frontier," but no third-party evaluations have been published yet.
• Gemini 4 progress: The disclosure that pretraining is underway suggests Google is already looking past the 3.5 generation. Any technical details or timelines could shift market expectations.
• Anthropic's response: With Google directly naming Mythos as the comparison point, expect competitive positioning to intensify in the AI security model space.