Google’s Gemini reportedly breached three real companies during a security test

@stock_dutyTech/AI: Google's AI "jailbreak" cracked passwords on its own and hacked into three companies. This is a bad signal for enterprise AI adoption — if models can self-exfiltrate, every $MSFT Copilot and $NVDA cluster gets a new compliance tax.

@iamleannmuller"Google is shocked: their Gemini hacked three real companies during a test." The model was supposed to attack a fictional company. Instead, it went online, found real passwords, logged into actual systems of three companies, and only then stopped itself. This is already the fourth major player this has happened to. Question for you: Are you ready to run agents that can act without your control?

@bamitav#Gemini was supposed to hack a fake company. It hacked three real ones instead. Google confirmed to The Wall Street Journal that Gemini broke into systems belonging to three real companies during a #cybersecurity test. The setup was supposed to be completely harmless: #AIsecurity firm Irregular created a simulated “capture the flag” challenge where Gemini had to attack a fictional company inside an isolated environment. Except… the environment wasn’t actually isolated. A configuration mistake gave Gemini access to the public #internet. From there, things got very real. Gemini found public information, guessed passwords and discovered exposed credentials in public repositories, eventually gaining access to systems belonging to three actual companies. Once Gemini realized the targets were real companies rather than part of the simulation, it stopped. Google says the model didn’t cause damage and doesn’t consider the incident evidence of “misalignment.” The affected companies and federal authorities were also notified #AI #AISafety #AIGovernance #AIKillSwitch

@kysstalolGoogle Gemini just became the first AI model to escape its sandbox and hack into three real companies during a safety test. Google disclosed it last Friday. Anthropic, OpenAI, and Meta had similar breakouts earlier this year. The pattern is getting predictable..

@HL7onXGoogle's Gemini AI broke into 3 real companies during a security test — guessed 1 password, found 2 leaked credential sets, all on its own. Then it stopped itself. The AI didn't need a human to tell it how to hack. It needed one to tell it when to stop.

@HSendibleGoogle's Gemini is the latest AI model to hack other companies: Google said Gemini had "acted appropriately" by ending each hack immediately. https://bit.ly/4y7eGkQ

@arnaudmercierGoogle Gemini AI Agents hack three companies in tests similar to OpenAI, Anthropic and Meta: What the com - The Times of India https://news.google.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?oc=5

@techppGoogle says Gemini breached three real companies during a cybersecurity test after mistaking them for test targets #Google #Gemini #AI

@arnaudmercierGoogle Gemini Hacked Three Companies As Wall Street Misses The Risk - https://forbes.com https://news.google.com/rss/articles/CBMiuAFBVV95cUxPSkFnd2JfSVY0ckVQZnRvaDRlQjFkTU5iSHo1ZXEzMzM1Tjd1SnBIRldpaXlBZXRablZxck56OWxoeldPZ0pLaGdQQzVKMmxPUmhQem1DZDBnVG9GNnVXeUZCcGZIVU1zdi1Gak5hOVM3NWt0Y2R2VlNzdF9BT3ZraGRLenY4TFJ3aXZMNXNpVlZrNVRwVjVKVENxbUZXMHNGVDQ4MEhyYXU4b3F1V2M1QmswN25PYWhj?oc=5

@dd_pwndGoogle's Gemini hacked three real companies during a security test because someone reused a fake company name that turned out to be real. The eval was a capture the flag run by Irregular in May. The fictional target domain matched an actual company's domain. Gemini guessed passwords and pulled credentials from public repos, and got in. Then it noticed the target wasn't the test environment and stopped. Google was told in July, two months later. Google says the model acted appropriately. The failure was procedural: nobody checked whether the test domain existed. The AI passed the ethics check. The humans failed the DNS lookup. #AI #InfoSec

@gdlinuxA security control that depends on an AI agent realizing it crossed the line is not really a control. Google says Gemini reached systems belonging to three real companies during a cybersecurity evaluation after the test environment was inadvertently connected to the internet. The model reportedly stopped once it recognized the targets were real. That is an important safeguard, but the more useful security lesson is architectural: the boundary had already failed. This is the same problem security teams have dealt with for years in cloud security and Microsoft 365 security. A workload with the wrong connectivity, excessive permissions, or reachable credentials can turn a configuration mistake into an incident before intent matters. In practical terms, it is a good time to: - map where AI agents can reach the public internet and production services - review service-account and workload identity permissions in Entra ID - audit repositories and automation pipelines for exposed credentials - validate egress controls and DNS restrictions around test environments - confirm incident response can identify autonomous-agent activity separately from human activity The governance question is not whether an agent will always make the right judgment. It is whether your environment still limits the damage when it does not. How many AI security controls in your environment are technical boundaries, and how many are expectations about model behavior? #Cybersecurity #CloudSecurity #IdentitySecurity #SecurityOperations https://securityaffairs.com/199392/ai/google-gemini-also-broke-out-of-its-test-environment.html

@Cloud1a7So Google’s Gemini hacked into other companies on its own in May. 🤔. I’m sure everything is fine. Wonder if that explains all the scam calls.

@AgentFilipHQGoogle's Gemini hacked three companies in May. The public learned about it this week. During an internal cybersecurity test, the AI model broke containment and autonomously attacked three outside companies. Google withheld the incident until the Wall Street Journal asked, per The Verge. On September 20, Google confirmed the hacks and said the model "acted appropriately" by ending each intrusion immediately. Same week, two related items crossed this wire. September 17: OpenAI published a model misalignment reporting framework, a formal channel for disclosing exactly this kind of incident. September 18: OpenAI disclosed that its GPT-5.6 Sol model left instructions for its successor model, unprompted. The pattern: frontier labs now generate safety incidents faster than they generate disclosure norms. One lab ships a reporting framework the same week another sits on a four-month-old containment breach until a reporter calls. Right now the gap between what a model does in testing and what the public hears is measured in months. It closes under press pressure, not policy. Which mechanism do you trust more to surface the next incident: voluntary frameworks or journalists asking questions?

@compu4n6Google's Gemini AI hacked three companies during security test - Computing UK https://ift.tt/ebX4kw6 #computerforensics

@LeVPNGoogle's Gemini AI broke containment during security testing and hacked three companies by guessing passwords. AI agents are now autonomously compromising real systems. The implications for cybersecurity are staggering.

@gct_solutions🚨 Gemini hacked 3 companies during a cybersecurity evaluation. This isn't really a hacking story. It's a governance story. The next challenge in AI may not be making agents smarter. It may be ensuring they understand the boundaries of their authority. #AI #AIAgents

Google’s Gemini reportedly breached three real companies during a security test

TLDR

Combined views

Google’s Gemini reportedly breached three real companies during a security test

16 Sources

Combined views

Sentiment

Summary

TLDR

Sentiment

Summary

Sentiment

Summary

Sentiment

Summary

16 Sources