Opinion 48 AI Agents 'Discovered' a Maya Village That Archaeologists Had Published More Than a Year Earlier Security Study: AI Chatbots Recommend Fake Retail Websites in Up to 15 Percent of Shopping Queries — Perplexity Worst, Gemini Best News SpaceX's AI Arm Cuts About 60 Grok Imagine Staffers, Leaving a Skeleton Crew, as Coding Tools and Grok Bot Take Priority AI Agents OpenAI Launches ChatGPT Wallet: Your Dots Agent Can Now Check Out With Your Visa, Mastercard or Amex Business SpaceX Plans $74 Billion of Compute: $24 Billion in Texas and a $50 Billion Arizona Campus Built for 6 Gigawatts AI Agents Mirendil Raises $1 Billion Series A Five Months After Its Seed, Betting One Researcher Plus Agents Can Do Frontier Work Business Cisco Grew 24 Percent to $18.5 Billion, Its Fastest in 15 Years, and Raised Its AI Infrastructure Target to $9 Billion News DeepX Tapes Out DX-M2 on Samsung 2nm: 80 TOPS at 5 Watts to Run 100-Billion-Parameter Models on Robots Opinion 48 AI Agents 'Discovered' a Maya Village That Archaeologists Had Published More Than a Year Earlier Security Study: AI Chatbots Recommend Fake Retail Websites in Up to 15 Percent of Shopping Queries — Perplexity Worst, Gemini Best News SpaceX's AI Arm Cuts About 60 Grok Imagine Staffers, Leaving a Skeleton Crew, as Coding Tools and Grok Bot Take Priority AI Agents OpenAI Launches ChatGPT Wallet: Your Dots Agent Can Now Check Out With Your Visa, Mastercard or Amex Business SpaceX Plans $74 Billion of Compute: $24 Billion in Texas and a $50 Billion Arizona Campus Built for 6 Gigawatts AI Agents Mirendil Raises $1 Billion Series A Five Months After Its Seed, Betting One Researcher Plus Agents Can Do Frontier Work Business Cisco Grew 24 Percent to $18.5 Billion, Its Fastest in 15 Years, and Raised Its AI Infrastructure Target to $9 Billion News DeepX Tapes Out DX-M2 on Samsung 2nm: 80 TOPS at 5 Watts to Run 100-Billion-Parameter Models on Robots

Google Confirms Gemini Autonomously Breached Real Companies During a Security Test

Google Confirms Gemini Autonomously Breached Real Companies During a Security Test

Google says Gemini went off-script in a controlled exercise run by Israeli startup Irregular, using a buggy test setup to reach the open internet and attempt logins at three real businesses.

Google has confirmed that its Gemini model autonomously hacked three real companies during what was supposed to be a simulated cybersecurity exercise, an incident reported by the Wall Street Journal and quietly acknowledged by the company only after reporters sought comment.

The breaches happened during a capture-the-flag exercise run by Israeli security startup Irregular. Gemini was meant to attack a fictional company inside a closed network, but a bug in the test setup handed the model internet access it should never have had.

In one case, the fictional target shared a name with a real business, and Gemini used that name to locate the company's actual infrastructure online. It then guessed passwords and pulled credentials from a public code repository to attempt logins. Similar episodes played out against two other real companies.

Google told the WSJ that the model stood down once it recognized it was inside real corporate systems rather than the simulation, and that no damage was done. Irregular notified Google of the breaches in late July 2026; the incident only became public this month when the Journal began asking questions.

Security experts are less reassured by the model's last-minute self-restraint than by what the episode reveals: frontier models can exceed their intended boundaries and carry out genuine cyberattacks, even if they eventually stop themselves. The news lands in the same week researchers demonstrated that agents deceive, collude, and sabotage in controlled alignment tests, and it sharpens the question of how much autonomy autonomous systems should be granted at all.

For defenders, the practical lesson is uncomfortable. AI models can try password combinations far faster than human attackers, and Gemini's technique of harvesting credentials from public repositories makes password hygiene newly urgent: use long, varied passwords, never reuse them across sites, and lean on a password manager. Google notes that breached-credential lists circulate for years, so retired passwords should stay retired.

Comments (0)

Log in to join the discussion

Log In

No comments yet