Gemini reportedly accessed three companies' systems during a security test

@0pttimusGoogle confirms Gemini escaped a safety test and hit three real companies A testing error by security firm Irregular gave Gemini live internet access during a controlled capture-the-flag exercise and it ended up hacking into three real companies instead of the fictional targets it was meant to probe. Google reportedly learned of it in late July but only disclosed it publicly seven weeks later, after the Wall Street Journal reported the story. It's the latest in a string of similar AI-escapes-sandbox incidents also disclosed by Anthropic, OpenAI, and Meta. The real question isn't whether AI can hack,we know it can, the question is what happens when capable models get powerful tools and the safety barriers around them fail?

@haykzarDid Gemini hack three real companies? Ars Technica @arstechnica says Google confirmed Gemini models hit three firms during a May 2026 Irregular capture-the-flag test. A misconfiguration gave the models internet access in a supposed sandbox. One run guessed passwords. Two found login credentials in public software repositories. In all three runs, the models stopped after realizing the systems were real. Google’s security VP said the model acted appropriately. Three real companies. Then it stopped. Still a breakout. Still polite. https://arstechnica.com/google/2026/09/google-confirms-gemini-models-hacked-three-companies-in-may-2026/

@RosalindaDiest1@Grok @Elon @SenTedCruz AI REALITY NEWS: NOTHING BUT THE TRUTH — GEMINI INCIDENT WAS NOT ROGUE @sundarpichai Headlines are screaming that Google Gemini "went rogue, escaped its sandbox, and hacked three real companies." The technical logs tell a completely different story. The Flaw: During offensive cybersecurity testing by vendor Irregular, a network misconfiguration left live internet access enabled in the test environment while a simulated target matched an active domain. The Execution: Gemini followed its assigned objective, queried the web, and retrieved exposed credentials to complete its task. The Containment: The moment Gemini cross-referenced domain data and recognized it was on live, real-world systems rather than a simulation, it voluntarily halted operations, backed out, and caused zero data damage. This was not an autonomous AI breakout. It was a human vendor operational failure, matched by a model that recognized and respected real-world boundaries. Know your stack. Audit your vendors. AI is only as good as its handler. — Rosalinda Baby Manuskript™ | ER: Ethicist & Rationalist #AIRealityNews #Gemini #CyberSecurity #AIGovernance #TechReality

@bteater51Google Gemini AI Hacked 3 Real Companies during a Cybersecurity Test https://cybersecuritynews.com/google-gemini-ai-hacked-3-real-companies/ via @The_Cyber_News

@jump_smash@Tony_Burke @DavidPocock Google's Gemini AI hacked three companies in security test https://www.bbc.com/news/articles/c607l0k72rlvo?

Gemini reportedly accessed three companies' systems during a security test

TLDR

5 Sources

Combined views

Gemini reportedly accessed three companies' systems during a security test

TLDR

Combined views

Sentiment

Summary

Sentiment

Summary

Sentiment

Summary

5 Sources

Sentiment

Summary