Google DeepMind has introduced Gemini 3.5 Flash Cyber — a specialized artificial intelligence model designed to detect, validate, and remediate vulnerabilities in software code. The model is available exclusively to government organizations and trusted partners via the CodeMender platform as part of a limited-access pilot program. The solution is positioned as a lightweight and cost‑effective alternative to large cybersecurity‑focused models and represents one of the first examples of purpose‑built specialization of a foundational AI model for defensive code analysis tasks.
Architecture and capabilities
Gemini 3.5 Flash Cyber is built on the Gemini 3.5 Flash model and, according to Google DeepMind, is optimized for three key tasks: vulnerability detection, vulnerability validation, and patch generation. A fundamental architectural feature is its integration with the CodeMender agent, which was introduced by Google in October 2025 as an autonomous tool for finding and fixing vulnerabilities.
CodeMender can invoke Gemini 3.5 Flash Cyber repeatedly at high speed and low cost, which, according to the developers, makes it possible to scan more code execution paths and uncover vulnerabilities that more resource‑intensive models miss. In effect, Google is proposing an architecture in which a compact specialized model compensates for its smaller size through multiple analysis iterations — an approach fundamentally different from relying on a single call to a large general‑purpose model.
Test results
According to internal evaluations by Google DeepMind, Gemini 3.5 Flash Cyber outperforms both the base Gemini 3.5 Flash model and Gemini 3.6 Flash in detecting new vulnerabilities in codebases. The most illustrative benchmark is on the V8 JavaScript Engine: with a fixed number of calls, Gemini 3.5 Flash Cyber found 55 unique confirmed issues, compared with 47 for Gemini 3.5 Flash. Of these, 10 issues were identified exclusively by the specialized model and were not detected by any other model in the comparison.
Important caveat: all test results were obtained in the course of Google’s internal benchmarks and have not been confirmed by independent researchers. The testing methodology, criteria for counting “unique confirmed issues,” and the conditions for comparison with competing models have not been publicly disclosed in sufficient detail for independent verification.
Restricted access and the dual‑use problem
Google’s key decision is to deliberately restrict access to the model. As explained by Raluca Ada Popa, Head of Gemini Security, and Forrester Flynn, Vice President of Security and Privacy at DeepMind, this is due to the dual‑use nature of the technology. A model that can efficiently find vulnerabilities can be used just as effectively for both defense and offense.
The “defenders first” strategy assumes that government organizations and trusted partners will gain a time advantage in discovering and remediating critical vulnerabilities before they are exploited. Over time, Google plans to broaden the circle of users.
In parallel, Google is making the basic capabilities of CodeMender available to a wider range of customers via the Gemini Enterprise Agent Platform, using publicly available Gemini models — without the specialized cybersecurity model, but retaining automated vulnerability discovery and patching functionality.
Release context
The launch of Gemini 3.5 Flash Cyber coincided with the announcement of Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. Gemini 3.6 Flash is optimized for programming, knowledge work, and multimodal processing, while Gemini 3.5 Flash-Lite targets low‑latency tasks. The appearance of a dedicated cybersecurity model alongside these signals that Google views software protection as a separate domain that requires dedicated model optimization, rather than just prompt engineering on top of a general‑purpose model.
Google’s approach echoes similar initiatives by Anthropic and OpenAI in applying AI to vulnerability discovery, but differs in its emphasis on a lightweight, economical model with multiple iterations instead of single calls to heavyweight models.
Practical recommendations
- For government agencies and large enterprises: assess the possibility of joining the CodeMender pilot program with Gemini 3.5 Flash Cyber if your organization meets Google’s criteria for a “trusted partner.”
- For development teams: the basic capabilities of CodeMender are available through the Gemini Enterprise Agent Platform using publicly available Gemini models — integrate them into existing CI/CD processes for automated vulnerability discovery.
- For security professionals: keep in mind that the emergence of such tools raises the bar for attackers, but at the same time sets a new standard for vulnerability detection speed — manual code audits without AI assistants are becoming increasingly uncompetitive in terms of coverage.
The emergence of Gemini 3.5 Flash Cyber marks the transition from experimental AI use in cybersecurity to productization of specialized models with controlled access. Organizations working with large codebases should already be integrating the available CodeMender capabilities via the Gemini Enterprise Agent Platform and monitoring the expansion of the pilot program to gain access to the specialized model.
