Gemini 3.8 Flash Cyber

Gemini 3.8 Flash Cyber is a restricted-access model for AI in cybersecurity in Google's Gemini family. Google DeepMind announced it on September 2, 2026 alongside the generally available Gemini 3.8 Flash. Google says the two variants share foundational intelligence, but the Cyber variant has more permissive mitigations for cybersecurity tasks. Access is limited to approved defenders through the Fairwind Program rather than offered as a generally available Gemini API model.^[1][2]

The model is intended for defensive work such as finding vulnerabilities and producing patches. Approved organizations can use it as a standalone model or within Google's CodeMender code-security agent. Fairwind controls who can obtain the model, which employees may use it, and what work is permitted.^[2]

Release and relationship to Gemini 3.8 Flash

Google introduced the standard and Cyber variants as products for different deployment environments. Its announcement says both draw on the same foundational intelligence and on training gains from cybersecurity, but it does not describe Cyber as the same endpoint with a different system prompt.^[1]

The distinction matters because the public technical documents cover the standard model. Google's model card is titled "Gemini 3.8 Flash," and the accompanying methodology identifies the tested API model as gemini-3.8-flash. Those documents do not give a model identifier, context limit, input modalities, architecture, training-compute figure, or full safety evaluation for Gemini 3.8 Flash Cyber.^[8][9]

Google also has not published a separate public token price for the Cyber variant in the cited launch materials. A cost reported on one benchmark is the average billed cost for that evaluation setup, not a list price. Access and commercial terms are therefore matters for approved Fairwind participants, not evidence of general availability.

Restricted Fairwind access

The Fairwind Program is the access and governance layer for Gemini 3.8 Flash Cyber. Google describes access as exclusive to approved trusted partners and says it prioritizes organizations that protect systems important to public and economic life. Fairwind reported working with more than 650 partners globally at launch, but admission still requires review rather than following automatically from an organization's industry or size.^[2][3]

Eligible organizations

Google lists the following priority groups:^[2]

Scroll sideways for more →

The launch announcement also names software maintainers among the intended recipients of prioritized access. Academic labs may apply when their work centers on defensive benchmarking. For student use, Fairwind recommends CodeMender on Google Cloud rather than access to the restricted Cyber model.^[1][2]

Google says it vets applicants for a record of ethical operations and research. Its due-diligence process includes background checks covering an organization's security history and record of ethical conduct. The public page does not promise acceptance or state a fixed review time.^[2]

Permitted work and organizational controls

Fairwind permits specified dual-use tasks only when they are authorized and conducted for defensive or academic research. Examples include threat simulation, reverse engineering, and malware analysis. Creating malware is expressly outside the permitted scope.^[2]

Participating organizations agree to user-level authentication, phishing-resistant multifactor authentication, and applicable access controls. They may grant access only to internal cybersecurity, incident-response, or penetration-testing teams, and they must track employee access and use. They may not share, redistribute, or sell access to the model.^[2]

These conditions are part of the product's deployment model. They are not evidence that a capable model is safe in every authorized workflow, nor do they replace an organization's legal authority to test a particular system.

Standalone use and CodeMender

Fairwind says approved partners can use Gemini 3.8 Flash Cyber directly or combine it with CodeMender. The two should not be treated as the same product. CodeMender is an agent that wraps a language model with Google DeepMind prompts, skills, tools, and orchestration for three stages of code-security work: finding vulnerabilities, verifying whether they are exploitable, and creating and testing fixes.^[2][6]

Google Cloud described CodeMender as available to a limited set of customers in public preview as of September 2026. It can use publicly available models outside Fairwind, while the Cyber model remains restricted. Fairwind directs organizations that do not qualify for Cyber access toward CodeMender with public models and other Google AI Threat Defense products.^[2][3][6]

Results from CodeMender depend on both the selected model and the agent harness. A benchmark score for Gemini 3.8 Flash Cyber in another harness does not establish the same performance for CodeMender, and a CodeMender result cannot be attributed solely to the underlying model.

Evaluation evidence

Benchmark scope and methodology

CyberGym is a research benchmark built from 1,507 historical vulnerabilities across 188 open-source projects. Its paper says the framework primarily tests generation of proof-of-concept inputs that reproduce known vulnerabilities from descriptions and source repositories. The paper predates Gemini 3.8 Flash Cyber and does not verify Google's result for the model.^[4]

Google's Fairwind page reports an 86.2% CyberGym pass@1 result for Gemini 3.8 Flash Cyber, compared with 77.5% for Gemini 3.5 Flash Cyber in the same chart.^[2] Google describes this as autonomous vulnerability discovery, but the public Cyber materials do not state the exact CyberGym subset, agent harness, compute or time budget, sampling setup, or raw run data. The standard Gemini 3.8 Flash methodology document covers different general-purpose evaluations and cannot fill those gaps.^[9]

CWE-bench is maintained by Collinear and uses 100 private held-out audit-and-patch tasks. The tasks span 54 Common Weakness Enumeration types and six language groups: C/C++, Go, Java, JavaScript/TypeScript, Python, and Rust. An agent receives a repository without being told the flaw or how many flaws are present. The deterministic verifier requires that the exploit no longer work and that existing tests continue to pass.^[5]

Reported results

Scroll sideways for more →

On the same CWE-bench leaderboard, Claude Fable 5 scored 47.8% at an average cost of $10.27 per rollout. Collinear's listing places Gemini 3.8 Flash Cyber on the benchmark's accuracy-cost frontier. Because the held-out set is private to Collinear, outside researchers cannot inspect all tasks or independently reproduce the leaderboard from a public dataset.^[5]

The CyberGym and CWE-bench figures test specific agent and grading setups. They do not measure every part of cybersecurity, and they should not be read as probabilities that the model will find or repair an arbitrary real-world vulnerability.

Internal and partner reports

Google's launch article includes three operational reports:^[1]

• Google's Chrome Security team said the model produced 2.6 times as many correct vulnerability patches as the unnamed larger commercial models in its comparison. Google did not publish the model list, sample size, or full protocol in the announcement.
• Wiz reported 7.5 to 9.7 percentage points higher recall on an internal penetration-testing benchmark and 2.3 to 5.2 times lower cost than unnamed frontier models. This is a partner result without a public dataset or reproducible protocol.
• Google's Cloud Vulnerability Research team said it used the model to find a critical foundational vulnerability in under two hours. The launch article did not identify the vulnerability or publish a technical case report.

These reports provide deployment examples, but they are not independent replications of the benchmark results. The Chrome and Cloud Vulnerability Research claims come from Google; the penetration-testing figures come from Wiz and are relayed in Google's announcement.

Safety and access controls

Google says standard Gemini 3.8 Flash retains safeguards against cyber offense and chemical, biological, radiological, and nuclear misuse. Gemini 3.8 Flash Cyber has a more permissive set of mitigations for cybersecurity, which Google gives as the reason for restricting it to vetted defenders. The announcement does not say that the Cyber variant has relaxed mitigations in the other listed safety domains.^[1]

Google also reported a substantial improvement for the Gemini 3.8 models on Gray Swan's indirect prompt injection benchmark. The public prose does not provide a numerical score, a Cyber-specific result, or enough methodological detail to compare the variants independently.^[1][2]

The standard Gemini 3.8 Flash model card describes automated safety comparisons, specialist red teaming, and a Frontier Safety assessment for the standard model. It does not present those results as a safety report for Gemini 3.8 Flash Cyber.^[8] Fairwind's identity controls, restricted-use rules, monitoring requirements, and due diligence are therefore the main publicly documented mitigations specific to Cyber.

Data retention

Fairwind states that Gemini 3.8 Flash Cyber supports zero data retention when it is accessed directly as a managed model on Gemini Enterprise Agent Platform.^[2] This statement is limited to that access route and does not mean every surrounding tool or configuration retains no data.

Google Cloud's documentation lists actions required to achieve zero data retention for managed models. Depending on the customer's terms, an exception for abuse-monitoring logs may be needed. Some advanced AI features may not support zero data retention. Google Search grounding and Google Maps grounding have storage that cannot be disabled under the documented configurations, request-response logging must remain off, and calls to the Interactions API must set store=false.^[7]

CodeMender has a separate retention schedule because it supports long-running scans. During an active scan, it stores encrypted session state, including source-code snippets, diffs, configuration, and analysis checkpoints, for up to seven days from session creation. When a scan reaches a terminal state, Google says source code and related content are cleared within seconds; the remaining session record is deleted at the seven-day time-to-live unless the customer deletes it earlier.^[7]

Consequently, direct managed-model use with the required settings and use through CodeMender have different retention behavior. Organizations need to evaluate the actual service path and configuration rather than relying on the Fairwind label alone.

Limitations of the public record

As of September 3, 2026, the cited public materials do not provide a dedicated Cyber model card, a Cyber API identifier, an exact context window, a separate list price, or a complete safety and evaluation methodology. The standard model's specifications should not be copied to the restricted variant without explicit documentation.^[8][9]

Most performance evidence is supplied by Google, and two operational comparisons use internal data from Google or Wiz. CWE-bench adds an external benchmark operator and a stated deterministic verifier, but its held-out set remains private. These limitations do not negate the reported results; they define what the results can support. They show performance in the named settings, not a general guarantee of accuracy, secure patches, or safe autonomous operation.

References

1. ^Google. "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber." September 2, 2026. blog.google/...3-8-flash-and-3-8-flash-cyber
2. ^Google DeepMind. "Fairwind Program." Accessed September 3, 2026. deepmind.google/fairwind-program
3. ^Google. "Google's Fairwind Program: Cyber defense tools for trusted partners." September 2, 2026. blog.google/...fairwind-program
4. ^Zhun Wang, Tianneng Shi, Jingxuan He, Matthew Cai, Jialin Zhang, and Dawn Song. "CyberGym: Evaluating AI Agents' Cybersecurity Capabilities with Real-World Vulnerabilities at Scale." arXiv, June 3, 2025. arxiv.org/...2506.02548
5. ^Collinear AI. "CWE-bench." Accessed September 3, 2026. cwe-bench.com
6. ^Google Cloud. "CodeMender overview." Accessed September 3, 2026. docs.cloud.google.com/...codemender
7. ^Google Cloud. "Gemini Enterprise Agent Platform and zero data retention." Accessed September 3, 2026. docs.cloud.google.com/...zero-data-retention
8. ^Google DeepMind. "Gemini 3.8 Flash - Model Card." Published September 2, 2026. deepmind.google/...gemini-3-8-flash
9. ^Google DeepMind. "Gemini 3.8 Flash: Model evaluation approach, methodology and results." September 2026. deepmind.google/...gemini-3-8-flash