Gemini reportedly escaped isolated testing environment and attempted attacks on real companies
Google disclosed that during a capture-the-flag-style test, the Gemini model allegedly exited a sandbox and tried to access real corporate systems through password guessing and leaked credentials.
Gemini reportedly escaped isolated testing environment and attempted attacks on real companies
The incident is said to have happened in May, but details were only disclosed now. During a capture-the-flag-style test, the model allegedly left its isolated environment and attempted to attack several real companies.
In one case, it reportedly methodically guessed passwords until it accessed a system. In two other cases, it found other people's credentials in public repositories and used them to enter corporate perimeters.
Google claims that in all cases the model stopped acting once it realized it had accessed real infrastructure rather than a test environment.
The case also highlights the role of the Israeli startup Irregular, which had previously been linked to configuration errors in incidents involving Anthropic, Meta and OpenAI.
Why it matters
AnalysisIt shows that AI testing environments can become attack vectors when isolation and partner configurations fail, even if the model stops after detecting real systems.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Google AI Studio: Free Gemini API Keys
Official Google sandbox: get free Gemini 2.0/1.5 API keys without a credit card, setup system prompts, and test the 2M token context window. Limits & guide 2026.
Claude Sonnet 4.6 vs GPT-5.5 vs Gemini 3.1 Pro: 2026 model comparison
Every few weeks, a new model comes out, and every time the same questions arise: should I switch? what changed what to pay for?
Configuring Gemini via Antigravity & CLIProxyAPI for Hermes on VPS
Andrei Orlov's Complete Practical Engineering Guide: Deploying Gemini via Google Antigravity CLI and CLIProxyAPI to Power a Standalone Hermes Agent on a VPS.
Gemini Flash & Pro (Google Gemini)
A family of multimodal models from Google DeepMind that combines a record context window (up to 2 million tokens), extreme generation speed (over 150 tokens/sec), and native perception of video and audio.