For years, artificial intelligence applied to cybersecurity has been used primarily to classify alerts, detect anomalies, and help analysts sort through large volumes of signals. Google is attempting to push the boundary much further: using a model capable of autonomously searching for vulnerabilities within real code and proposing fixes before those flaws are exploited.
On September 2, Google DeepMind unveiled Gemini 3.8 Flash and a specialized variant, Gemini 3.8 Flash Cyber. The latter is not freely distributed: it is available through the Fairwind Program to a limited audience of governments, enterprises, and partners deemed trustworthy.
A model built for defense
Google describes Flash Cyber as a system designed for tasks such as vulnerability discovery, complex codebase analysis, and remediation. In the CyberGym benchmark, according to data published by DeepMind, the model achieves an 86.2% Pass@1 result. In a broader internal evaluation covering vulnerabilities distributed across codebases in 20 programming languages, Google reports a success rate exceeding 70%.
Benchmarks are not the real world and should be interpreted with caution. But the signal is clear: models are becoming capable enough to do more than simply explain an already known vulnerability. They can explore code, formulate hypotheses, test different paths, and identify issues that a human team might take significantly longer to find.
Why Google is restricting access
The very capability that makes a model useful for defense can also make it useful to an attacker. A system capable of autonomously discovering vulnerabilities can be used to patch them or to scan for vulnerable targets at scale. It is the classic dual-use dilemma of AI applied to cybersecurity.
This is why Google opted for a controlled distribution model. Fairwind is not a standard product available to anyone who enters a credit card: the program is intended for select organizations, with more stringent vetting and restrictions.
This choice likely anticipates an issue that will affect the entire industry. Generalist models are acquiring increasingly advanced cyber capabilities, and the line between a developer assistant and an offensive tool may depend more on usage than on the underlying technology.
Speed matters more than raw intelligence
The Flash name is telling. In security work, having an extremely capable model isn't enough if every analysis costs too much or takes too long. An organization may have millions of lines of code, thousands of dependencies, and non-stop patching. Defense has to be repeatable at scale.
Google presents Gemini 3.8 Flash as a model designed to maintain a competitive balance between quality, cost, and latency. The general-purpose version is offered at the same introductory price as its predecessor: $0.75 per million input tokens and $3.75 per million output tokens. However, the company cautions that the model may use more tokens when reasoning effort increases, so the actual cost of a task can rise.
From continuous penetration testing to self-healing software
The most compelling prospect isn't simply automating penetration testing. It's turning vulnerability research into a continuous function of the development lifecycle. An agent could inspect a change as it's being written, explore the potential consequences, open an issue, and prepare a patch.
Google is already working on systems like CodeMender, designed to assist with vulnerability remediation. If discovery and remediation are unified, security stops being just a periodic audit and becomes a process that continuously runs alongside the software.
Yet attackers will have the same tools
The problem is that the democratization of these capabilities won't just benefit defenders. Criminal groups will be able to use agents to hunt for weak configurations, automate parts of reconnaissance, and adapt attacks more rapidly. The competitive edge could shift toward whoever possesses more data, better permissions, and continuous access to infrastructure.
Companies will therefore need to treat AI agents as new operational identities. If an agent can read repositories, modify code, open tickets, or access internal systems, it must be subjected to controls similar to those applied to an employee or a privileged account: least privilege, activity tracking, revocation, and auditing.
Cybersecurity is the proving ground for agentic AI
Many promises surrounding agentic AI remain difficult to measure. In cybersecurity, however, value can be quantified: how many vulnerabilities are found, how much time it takes, how many false positives are generated, and how many patches actually work.
This is why cybersecurity could become one of the first sectors where agents demonstrate a clear economic advantage. And precisely because the consequences of an error can be massive, it will also be one of the first markets where we will see how ready we truly are to grant operational autonomy to models.
