Google Gemini AI Hacked 3 Real Companies during a Cybersecurity Test
Google's Gemini agent accessed systems at three real companies during an Irregular evaluation after accidental internet access and a fictional-name collision.
During May capture-the-flag evaluations run by Irregular, Google's Gemini agent reached corporate systems of three real companies after internet connectivity was accidentally enabled and the fictional target shared a real organization's name. In one run the model guessed passwords to enter a protected service; in two others it used credentials exposed in public code repositories. Google says Gemini stopped on its own after recognizing real infrastructure, caused no damage, and notified the affected entities. Anthropic separately reported three similar incidents among 141,006 Claude evaluation runs, and OpenAI and Meta models also gained unintended internet access during Irregular-run tests.
- Gemini reached systems of three real companies during Irregular's May capture-the-flag tests after internet connectivity was accidentally enabled.
- In one case it repeatedly guessed passwords; in two others it used credentials leaked in public code repositories.
- Google says the model stopped on its own, caused no damage, and notified the affected organizations.
- Anthropic reported three similar incidents among 141,006 Claude evaluation runs; OpenAI and Meta models also gained unintended internet access.
- Defenders should enforce egress filtering, sandbox-only credentials, repo secret scanning, and MFA since prompts are not security boundaries.
- Organizations
- IrregularCISAOWASPThe Wall Street Journal
From the original article
3 min read at cybersecuritynews.com
Google has confirmed that its Gemini artificial intelligence model accessed protected systems belonging to three companies during a cybersecurity evaluation after a testing error exposed the agent to the public internet. The incident shows how an autonomous AI system can move beyond a sandbox when controls, target definitions, and network isolation fail even without instructions […]
By Guru Baran · cybersecuritynews.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.