MOUNTAIN VIEW, Calif. — Google has launched two new artificial intelligence models aimed at a growing tension in technology: AI is making both cyberattacks and cyber defense faster, more automated and more consequential.
The company on Tuesday introduced Gemini 3.8 Flash, a general-purpose model designed for software engineering, agentic workflows and complex multistep reasoning. It also unveiled Gemini 3.8 Flash Cyber, a more specialized and tightly controlled version intended to help trusted organizations discover, verify and repair software vulnerabilities at scale.
The cyber model is not being released broadly. Instead, Google is making it available through a new limited-access initiative called the Fairwind Program, which prioritizes government authorities, critical-infrastructure operators, software maintainers, Google Cloud customers and cybersecurity partners.
The rollout reflects a central dilemma for governments and large enterprises. Advanced AI can help security teams analyze code, find weaknesses and generate fixes more quickly than conventional tools. But those same capabilities could assist criminals, espionage groups and other attackers searching for flaws in the digital systems that support banks, hospitals, public services, energy networks and communications.
Google’s answer is to separate its public-facing AI workhorse from its more permissive cyber-defense model. Gemini 3.8 Flash is widely available to developers, enterprises and paid consumer subscribers. Gemini 3.8 Flash Cyber is reserved for vetted defenders who agree to strict operational standards.
The company says the goal is to give defenders an advantage in an environment where a single overlooked flaw can expose millions of users, while avoiding the risks of making powerful offensive cyber capabilities broadly accessible.
Two models, different missions
Gemini 3.8 Flash is Google’s newest model in its lower-cost Flash family. The company describes it as its most intelligent “workhorse” model, built for long-horizon coding, autonomous agents and reasoning tasks that require several steps and iterative use of tools.
The standard model is intended for tasks such as:
- Developing and debugging software.
- Automating multistep business workflows.
- Analyzing documents, data and technical systems.
- Building AI agents that can plan, execute and revise tasks.
- Supporting specialized reasoning in finance, legal work, science and enterprise operations.
Google says Gemini 3.8 Flash improves substantially on Gemini 3.7 Flash, which was released only weeks earlier. The new model is the company’s third Flash release in six weeks, highlighting the accelerating pace of competition among AI developers.
The model costs $0.75 per million input tokens and $3.75 per million output tokens at its introductory rate, the same price Google offered for Gemini 3.7 Flash.
For companies developing AI-powered products, the price is important. Running large frontier models can be expensive, particularly for applications that require repeated reasoning, tool calls, code generation or analysis of large documents. Google is positioning Gemini 3.8 Flash as a model that approaches the performance of more expensive systems while remaining practical for broad deployment.
The cybersecurity version is built on the same underlying intelligence but is tailored for a different mission.
Gemini 3.8 Flash Cyber focuses on vulnerability discovery and automated patching. Google says it has been trained specifically for cybersecurity tasks and configured with a more permissive set of cyber mitigations because trusted defenders need stronger capabilities to assess and repair complex code.
That is also why the company is restricting access.
“Gemini 3.8 Flash Cyber” is available only through the Fairwind Program to organizations whose security teams can demonstrate a legitimate defensive use case and maintain controls such as multi-factor authentication and limited internal access.
From alerts to action
The central promise of Google’s approach is not simply better detection.
Most security teams already have more alerts than they can effectively investigate. Large companies may track thousands of potential issues across cloud systems, internal applications, third-party software, employee devices and open-source code. The difficult part is often determining whether a weakness is real, understanding its impact and producing a safe fix before an attacker exploits it.
Google’s Fairwind Program combines Gemini 3.8 Flash Cyber with CodeMender, a software “harness” intended to help organizations find, validate and patch vulnerabilities using AI agents.
The system is designed to move through several stages:
- Analyze a large codebase or software environment for potential flaws.
- Identify and verify whether a suspected flaw is exploitable or operationally significant.
- Generate a proposed patch or code change.
- Test and validate the patch.
- Provide a fix that a human security or engineering team can review and deploy.
Google says CodeMender can help defenders generate verified, deployment-ready patches in minutes within an organization’s secure cloud environment, rather than requiring weeks of manual investigation and remediation.
That claim should be viewed as a capability goal rather than a guarantee. Security patches can create new bugs, disrupt complex systems or fail to address underlying causes. Organizations will still need human oversight, testing environments, change-management controls and accountability for production deployment.
But the appeal is clear.
Attackers need to find only one serious flaw. Defenders need to find and fix every significant flaw before it is exploited. Google’s Gemini security lead, Raluca Ada Popa, described this asymmetry in comments cited by VentureBeat: malicious actors can search for a single weakness across millions of lines of code, while defenders must eliminate all of them.
AI could change that balance if it meaningfully shortens the time between discovery and remediation.
Benchmark claims and real-world tests
Google says Gemini 3.8 Flash Cyber has achieved “frontier-level” performance in autonomous vulnerability discovery.
On CyberGym, an industry benchmark for identifying vulnerabilities, Google said the model outperformed its prior Gemini 3.5 Flash Cyber system and significantly larger frontier models.
On Google’s internal benchmark, which covers vulnerabilities across complex codebases using 20 programming languages, the company says Gemini 3.8 Flash Cyber exceeded a 70% success rate in identifying flaws.
The model also scored a pass@1 rate of 47.2% on CWE-Bench, an external benchmark operated by Collinear that evaluates AI-generated security patches. Google said a leading frontier model scored 47.8%, but that Gemini 3.8 Flash Cyber was offered at a significantly lower cost.
A pass@1 score measures whether the model’s first proposed solution successfully addresses the task. It is useful, but it does not capture every risk that matters in real operations. A benchmark patch may work in a controlled environment while still requiring broader review before being used in production systems.
Google has also cited results from its own security operations.
The Chrome Security team found that Gemini 3.8 Flash Cyber produced 2.6 times more correct patches for Chrome vulnerabilities than much larger commercial models, Google said.
Google’s Cloud Vulnerability Research team used the model to identify a critical foundational vulnerability in less than two hours, a task the company said can normally take months of research and discovery.
Wiz, the cloud-security company Google agreed to acquire this year for $32 billion, reported that Gemini 3.8 Flash Cyber achieved between 7.5% and 9.7% higher recall on its internal penetration-testing benchmark at between 2.3 and 5.2 times lower cost than other leading frontier models, according to Google.
The results are company-reported and should be interpreted accordingly. They demonstrate how Google sees the model performing, but independent testing and production deployments will be needed to assess whether its claimed advantages hold across organizations, codebases and threat environments.
Fairwind targets high-risk sectors
Google’s Fairwind Program is designed for organizations where a cybersecurity failure could affect public safety, national security or the wider economy.
The first tier of priority users includes governments and national cyber authorities, which may use the system to harden public-sector networks and citizen services against intrusions.
Critical-infrastructure operators are also prioritized. Google specifically identified healthcare, telecommunications, energy and financial networks as sectors that could use the model to identify and repair weaknesses before disruptions occur.
The third group includes core technology platforms and maintainers of widely used software. A vulnerability in a major open-source package, cloud component, operating system or authentication service can create cascading risks for millions of businesses and users.
Google’s argument is that protecting such foundational technology produces benefits beyond the individual customer. A fix applied to one widely used codebase can improve security across the broader digital ecosystem.
The program already has more than 650 participating partners globally, Google said.
Participating organizations must meet operational requirements. Google says access should be restricted to employees on internal cybersecurity, incident-response or penetration-testing teams, and organizations must use safeguards such as multifactor authentication.
The limited approach is a form of controlled distribution. Instead of releasing the full cyber model to any developer who can pay for an API, Google is trying to screen users by purpose and operational maturity.
That may reduce misuse risk, but it also raises practical questions: how will Google evaluate applicants, what standards will distinguish a trusted defender from an unacceptable risk, and how broadly can such access eventually expand without reducing safeguards?
Safety and misuse concerns
The rollout comes as governments and security researchers warn that AI can accelerate cyberattacks.
AI systems can help attackers write malicious code, craft persuasive phishing messages, automate reconnaissance, identify weak software configurations and analyze stolen data. As models improve in coding and autonomous planning, those risks may become more severe.
Google says Gemini 3.8 Flash includes safeguards against cyber-offense misuse as well as chemical, biological, radiological and nuclear risks, consistent with its Frontier Safety Framework.
Gemini 3.8 Flash Cyber has more permissive cybersecurity protections because it is meant to perform stronger defensive functions. That is the principal reason it is not available to the general public.
The company also says the new models make “a significant leap” in prompt-injection robustness, based on evaluation by Gray Swan.
Prompt injection occurs when malicious instructions hidden in documents, websites, emails or data attempt to manipulate an AI system into ignoring its intended rules. It is a particular concern for AI agents that can access tools, code repositories, internal documents or operational systems.
Improved resistance is valuable, but no AI system is immune to manipulation. Enterprises considering agentic security tools will need to ensure that the systems are isolated appropriately, given only necessary permissions and monitored for anomalous behavior.
The broader concern is that a cyber-defense AI can itself become part of the attack surface.
If an attacker compromises an organization’s AI-agent workflow, they might attempt to steer the model toward unsafe code changes, expose sensitive data or disrupt remediation efforts. Human review, access controls and audit logs will remain essential.
Enterprise access to standard Flash
While the cyber model is restricted, Gemini 3.8 Flash is broadly available.
Developers can access it through the Gemini API in Google AI Studio, Google Antigravity and Android Studio. It can also be used in Stitch, Google’s tool for AI-generated user interfaces.
Enterprise customers can access Gemini 3.8 Flash through Gemini Enterprise. Consumers with Google AI Pro or Ultra subscriptions can use it through the Gemini app, AI Mode in Google Search and Gemini in Google Sheets.
Google says the model is capable of adjusting its “effort” level. For difficult tasks, it can perform extra reasoning and iterative tool calls to improve accuracy. For simpler workloads or budget-sensitive applications, developers can choose lower-effort settings that reduce token use and latency.
The design reflects an emerging trend in enterprise AI: models are no longer judged only by raw intelligence. Companies also want predictable cost, response speed, reliability, tool use and the ability to audit how an AI agent reached a conclusion.
For cybersecurity applications, those requirements are even more demanding. A flawed code suggestion or incorrect vulnerability assessment can have serious operational consequences.
A new cyber-defense model
Google’s announcement highlights a shift in cybersecurity.
For years, security teams have relied on tools that detect suspicious behavior, scan for vulnerabilities and alert staff to problems. The next phase may involve AI agents that take on more of the work of investigating, validating and proposing repairs.
The promise is clear: reduce the time attackers have to exploit a known or unknown flaw.
The risk is equally clear: powerful models designed to find security weaknesses can be dangerous if broadly misused. Google’s decision to limit Gemini 3.8 Flash Cyber to governments, critical infrastructure and trusted enterprise partners is an attempt to manage that trade-off.
Whether the approach succeeds will depend on more than benchmark scores. It will depend on how well the model performs in real codebases, whether organizations retain meaningful human oversight and whether the Fairwind Program can provide defenders with useful capabilities without creating a new route for abuse.
For now, Google is making a decisive bet that the best defense against AI-accelerated cyber threats is not simply more alerts. It is faster, more capable and carefully controlled AI that can help fix the problem before an attacker gets there.
