July 21, 2026 | Models
Raluca Ada Popa and Four Flynn
Google has invested in cybersecurity for years, pioneering automated vulnerability discovery to secure global codebases. Code security agents like CodeMender can automatically discover and patch critical software vulnerabilities. However, as the speed at which AI agents find vulnerabilities begins to outpace the speed at which defenders can patch them, addressing this global threat requires a highly capable, low-cost, and scalable approach.
Today, we are expanding our long-standing efforts to better equip defenders by introducing Gemini 3.5 Flash Cyber—a fine-tuned, lightweight cybersecurity model built on 3.5 Flash. It is designed to quickly and efficiently discover, verify, and patch vulnerabilities, performing these tasks more effectively than Gemini’s mainline Flash models.
The performance and efficiency of Flash make it the ideal foundation for our cybersecurity model work. Built on Flash, 3.5 Flash Cyber offers a lower-cost, highly capable alternative to large, expensive cybersecurity models.
Given the dual-use nature of this technology, we are taking a cautious approach to deploying 3.5 Flash Cyber. As part of a limited-access pilot program, 3.5 Flash Cyber will initially be available through CodeMender only to governments and trusted partners, with access expanding over time. This will give front-line defenders a head start, allowing them to find and fix critical vulnerabilities before they can be exploited, while mitigating the risk of broader abuse.
Additionally, we are making the core capabilities of CodeMender directly available to customers using general-purpose Gemini models through the Gemini Enterprise Agent Platform.
The Search Space Problem: The Advantages of Lightweight Models in Code Security
Discovering deep flaws requires exploring an extremely vast execution search space. Relying on a single, expensive call to a large model can create a bottleneck. 3.5 Flash Cyber is particularly well-suited for vulnerability discovery, where agents must scan large codebases and analyze numerous code paths.
CodeMender makes multiple calls to 3.5 Flash Cyber, enabling the agent to analyze far more code paths to discover and verify vulnerabilities. Subsequently, individual sub-agents generate a single, high-quality report.
Thanks to its speed and low cost, 3.5 Flash Cyber can be easily integrated into high-frequency scanning, time-sensitive release processes, or large-scale commit scanning pipelines.
3.5 Flash Cyber Benchmark Results: An Efficient Alternative Beyond Large Cybersecurity Models
We tested 3.5 Flash Cyber on several benchmarks. In particular, we evaluated it on the CyberGym benchmark, which assesses the ability of AI agents to handle hundreds of real-world software vulnerabilities. By allowing CodeMender to call 3.5 Flash Cyber up to five times when generating a final report—taking full advantage of its low cost—the overall agent achieved performance on CyberGym* comparable to significantly larger models.
CyberGym Success Rate (pass@1)
*Competitor results are based on self-reported vendor scores
We also stress-tested the model’s capabilities beyond CyberGym without safety guardrails. Google’s Big Sleep team independently built an evaluation framework focused on discovering critical, hard-to-find vulnerabilities in some of the world’s most complex codebases, such as Chrome and Safari. Here, 3.5 Flash Cyber significantly outperformed mainline 3.5 Flash and 3.6 Flash.
Big Sleep Evaluation Success Rate (pass@1)
3.5 Flash Cyber was also evaluated in Google Chrome’s production commit scanning pipeline. The vulnerabilities involved were not publicly disclosed, ensuring the benchmark remained “uncontaminated” for both Gemini and competitor models.
The results show a significant improvement for 3.5 Flash Cyber compared to 3.5 Flash. Note: More recent competitor model versions after Opus 4.6 were excluded from the comparison because their built-in safety guardrails caused them to refuse these tasks.
Chrome Production Commit Scanning Pipeline Success Rate (pass@1)
Furthermore, 3.5 Flash Cyber consistently discovered more unique vulnerabilities compared to mainline 3.5 Flash and Claude Opus 4.6. In tests with a fixed number of calls on the highly complex V8 JavaScript engine, 3.5 Flash Cyber identified 55 unique, confirmed issues, compared to 47 for mainline 3.5 Flash and 36 for Opus 4.6—including 10 issues that were missed by both of the other tested models.
Basic cybersecurity models sometimes get stuck in loops, repeatedly identifying the same issue while missing critical vulnerabilities. A robust model provides broader coverage, uncovering a larger number of unique issues.
As the number of calls increases, we find that 3.5 Flash Cyber continuously discovers new code paths and vulnerabilities.
Real-World Applications and Scaling Defense at Google
Benchmarks are only part of the story. 3.5 Flash Cyber within CodeMender is already finding and fixing vulnerabilities in Google’s internal codebases, including Chrome, Android, Cloud, Ads, and YouTube.
The faster discovery speeds enabled by lightweight models are already having a measurable impact.
For example, Google’s Cloud Vulnerability Research team used 3.5 Flash Cyber to proactively secure our systems faster than ever before. In just two hours, the model discovered a remote code execution (RCE) vulnerability in a public API and identified a memory corruption vulnerability in a sensitive production service. It then generated a 100% reliable RCE exploit, bypassing standard mitigation techniques such as Address Space Layout Randomization (ASLR) and Write XOR Execute (W^X).
Early feedback from Wiz and testers at Cloud CISO Security Engineering confirms that 3.5 Flash Cyber offers a significant boost in capability compared to the mainline 3.5 Flash model.
Empowering Defenders at Scale
Google’s leadership in software security gives us a unique advantage. For instance, the Google-run vulnerability database OSV.dev, which covers over 700,000 open-source vulnerabilities, along with more than a decade of OSS-Fuzz results, helps us identify the highest-quality vulnerabilities.
This allows us to move beyond synthetic cybersecurity examples and teach the model how real security professionals work. Our models have learned to operate industry-standard tools, read millions of lines of code in massive projects like Chromium, and independently handle complex security tasks that require hours of continuous, deep analysis.
By powering CodeMender with 3.5 Flash Cyber, we are building a highly capable, scalable, and affordable architecture designed to help more defenders secure their software.