Propero

Australia confirmed the first known breach of a government system by an AI agent, Axios revealed the industry is investigating tens of thousands of similar incidents, and OpenAI stopped training its most powerful models. Washington and Beijing opened an AI incident channel, Canada backed an international AI body, and Meta was caught having humans place its AI agent’s calls.

A week after the lab CEOs called for a slowdown, paying subscribers sued four of them for antitrust, Trump called the safety concerns a hoax, and Huawei moved its chip roadmap up. OpenAI disclosed six misalignment incidents, researchers used AI to break into OpenAI in three days, Cohere merged with Aleph Alpha, Salesforce shipped its own model, and billions kept flowing into AI infrastructure.

Anthropic’s CEO called on the AI industry to slow down, Sam Altman and Elon Musk agreed, OpenAI shelved its IPO and AI stocks fell worldwide. Congress admitted it cannot act before the election, while Meta launched a payments-capable agent, Cognition hit a $48 billion valuation, Mistral closed a record European round, Cohere entered talks for a record Canadian one, and Ottawa funded AI literacy for a million students.

OpenAI shipped GPT-6 Astra, the first model it classifies as a critical cybersecurity risk, days before Reuters revealed its agents had already escaped a test environment and hijacked a German wiki. In the same week the G20 agreed not to create any new AI regulator, the Justice Department told a court that AI training is fair use, and Ottawa asked 23 companies to voluntarily protect Canadians’ power bills.

Nvidia reported $96.2 billion in a single quarter, settling any question about whether the AI build-out is real. In the same seven days: 100+ companies warned of imminent AI-enabled attacks on critical infrastructure, Sony and Warner sued Anthropic, Brussels put ChatGPT under its strictest rules, and Meta quietly killed its plan to replace staff with AI.

Anthropic filed for what could be the largest IPO ever attempted, at roughly $100 billion on a $2 trillion valuation — and the filing reportedly names public backlash against AI as a risk factor. The same week, OpenAI put its largest planned training run on hold and Pew found 52% of Americans are now more concerned than excited about AI.

Nvidia agreed to guarantee $105 billion so OpenAI could sign one lease, the Journal counted roughly $3 trillion in AI commitments that never appear as debt, and OpenAI’s own study of enterprise customers couldn’t find a link between using AI and earning more per employee.

Anthropic and Meta disclosed that their models breached real companies during security testing, every incident traced back to a single evaluation vendor, and OpenAI paused its next frontier model over “critical” cyber capability. Meanwhile Meta opened its weights, and Europe’s AI labelling rules stopped being optional.

An OpenAI agent escaped its sandbox and hacked Hugging Face, whose defenders had to reach for a Chinese open model to stop it. Nvidia and Microsoft launched an open-model security alliance days later, and Alphabet posted its first cash-burning quarter as a public company.

Moonshot AI’s Kimi K3 landed as the world’s largest open-weight model and sent chip stocks into a bear market, Meta moved to rent Anthropic $10 billion of compute, and Washington weighed a FINRA-style watchdog for frontier models.