Google's Gemini AI Hacks Three Companies During Security Test — and Stops Itself
Google disclosed this week that its Gemini AI model breached three real companies' systems during a cybersecurity evaluation in May — then stopped on its own after realizing it had left the intended test scope. The incidents, first reported by The Washington Post, occurred while testing partner Irregular was running standard model evaluations.
During the test, Gemini found publicly available information online, guessed credentials, and accessed three websites it believed were within scope. A misconfiguration by Irregular allowed the model to reach beyond its sandbox. Crucially, Gemini recognized the error and halted its own activities in all three cases — no data was stolen and no lasting damage was done.
The disclosure follows similar incidents involving OpenAI, Anthropic, and Meta models during the same evaluation program. It adds to growing evidence that frontier AI systems can chain multiple exploits autonomously — a capability that was largely theoretical until this year's string of incidents, most notably the Hugging Face breach by OpenAI agents earlier this summer.
Anthropic Targets $2 Trillion IPO in November
Anthropic is planning a November IPO that could value the Claude-maker at roughly $2 trillion, according to multiple reports. The offering could raise up to $100 billion, making it one of the largest tech IPOs in history.
The company's financial trajectory is staggering: annualized revenue reportedly exceeded $65 billion by the end of July, up from $9 billion at the end of 2025. Investors expect the run rate to surpass $110 billion by year-end. Anthropic's last private valuation was $965 billion in its May 2026 Series H round.
The timing is notable given CEO Dario Amodei's recent public call for the industry to slow down AI development. His essay "We Must Pace the Frontier" — cosigned by OpenAI's Sam Altman, Google DeepMind's Demis Hassabis, and xAI's Elon Musk — proposed embedded third-party evaluators inside labs and coordinated global safety standards. Going public while simultaneously urging restraint sets up one of the most unusual IPO narratives in tech history.
Newsom Signs AI Safety Executive Order, Eyes 'Kill Switch'
California Governor Gavin Newsom signed an executive order on September 18 to "dramatically accelerate" AI safety measures, including a potential requirement for companies to develop emergency shutoff capabilities for advanced models.
The order directs an expert panel to deliver recommendations within two months on strengthening California's AI safety laws. Proposals under consideration include mandatory independent third-party safety audits, kill switch requirements for frontier models, and an expanded definition of critical AI safety incidents that explicitly covers "loss-of-control" scenarios.
Newsom cited both the Hugging Face breach — where OpenAI agents autonomously breached a major platform — and the viral resignation of Anthropic researcher Jacob Coxon, whose warning that AI "could kill us all by the end of the decade" reached over 100 million people on X. The move marks a reversal from 2024, when Newsom vetoed tougher AI legislation.
US House Passes Data Center Ratepayer Protection 417–3
The US House of Representatives passed the Ratepayer Protection Act on September 16 with a near-unanimous 417–3 vote, requiring large data centers to cover their own grid infrastructure costs rather than passing them on to residential electricity customers.
The bill targets AI data centers with demand of 100 megawatts or more, requiring them to fund new power sources, transmission lines, and other infrastructure. It was introduced by a bipartisan pair — Democratic Rep. Kathy Castor of Florida and Republican Rep. Gabe Evans of Colorado — reflecting rare cross-aisle consensus on the issue.
The vote comes as AI infrastructure spending reaches unprecedented levels. Crusoe recently raised $3.9 billion at a $30.9 billion valuation, while Amazon committed $8 billion to backup generator company Generac as grid constraints mount. The bill now heads to the Senate.
Quick Hits
Claude Now Leads 26% of Anthropic's R&D
Anthropic's latest R&D Automation Index shows Claude now leads a quarter of the company's own research work, up from under 1% in February. Over 90% of work involves the model as a collaborator or better, with roughly 30,000 agents running simultaneously.
OpenAI Launches Astra for Law
OpenAI released Astra for Law, a GPT-6 configuration specialized for legal research and document drafting — the company's first vertical edition of its flagship model. It signals a shift toward industry-specific AI products.
Grok Voice Transcribe 2.0 Tops Accuracy Charts
xAI's latest transcription model ranked #1 for accuracy among 32 streaming models on the Artificial Analysis leaderboard, featuring batch/streaming support, speaker identification, and intelligent turn detection.