Gemini 3.8 Flash Cyber Finds Snowflake RCE and Transit Admin Leak in Global Scan for Good Push
Google and Wiz announced the global expansion of Scan for Good on September 24, a programme that deploys Gemini 3.8 Flash Cyber against internet-facing systems belonging to hospitals, municipalities, transit operators, and core technology providers. The AI systems probe publicly exposed websites, APIs, and applications under authorisation, then route findings to human researchers for verification and remediation.
The initiative had been running for several months before the global announcement. The September 24 announcement cited two confirmed findings as programme examples:
Snowflake GitHub Actions RCE. Wiz’s Red Agent identified a script injection vulnerability in snowflakedb/snowflake-connector-net, Snowflake’s public open-source repository. The flaw was reported via Snowflake’s HackerOne vulnerability disclosure programme and patched on June 23, 2026; Wiz published a public research post on August 17, 2026. An unauthenticated user could execute arbitrary commands inside a GitHub Actions runner by opening a GitHub issue with a specially crafted title.
Transit operator admin session leak. At an unnamed public rail operator, the model found a production database exposing active administrator sessions. Exploitation of the leak would have allowed an attacker to take control of routes, schedules, service announcements, and administrator accounts across the entire transportation system. Wiz secured the system before the finding was disclosed.
Both vulnerabilities were autonomously discovered by the models. Human validation confirmed real-world impact in each case before any fixing was attempted.
The model and the programme
Gemini 3.8 Flash Cyber launched September 2 through Google’s Fairwind Programme, a controlled-access scheme that limits the model to government authorities, national cyber agencies, critical infrastructure operators, and core technology platforms. More than 650 organisations have joined globally. The US Cybersecurity and Infrastructure Security Agency (CISA) endorsed the initiative; acting director Nick Andersen called it a timely addition to defensive vulnerability discovery.
The model’s design makes the access gate unavoidable. A system capable of reading large codebases, identifying memory safety flaws, and reasoning about reachability is close enough to an exploit-writing system that Google explicitly cites dual-use risk as the reason for restricted distribution. Google says the Cyber variant carries more permissive cybersecurity mitigations than standard Gemini 3.8 Flash precisely because access is limited.
On CyberGym, a benchmark for autonomous vulnerability discovery, Google reports Gemini 3.8 Flash Cyber at 86.2% Pass@1 versus 77.5% for Gemini 3.5 Flash Cyber. Wiz’s internal penetration-testing benchmark shows 7.5 to 9.7 percentage points of higher recall at 2.3 to 5.2 times lower cost than leading alternatives. Both figures are vendor-reported.
Pricing and access
Gemini 3.8 Flash Cyber is billed at the standard rate of $1.50 per million input tokens and $7.50 per million output tokens. Standard Gemini 3.8 Flash has an introductory promotional rate of $0.75/$3.75 per million through December 31, 2026; the Cyber variant is excluded from that discount.
Standard Gemini 3.8 Flash is broadly available in AI Studio, Gemini Enterprise, and developer APIs. The Cyber variant has no public API or self-serve sign-up route. Access runs through Fairwind application only.
The Scan for Good programme itself has no stated end date. Google describes the scope as global and will add organisations as they apply.