Altman Says He's Open to Slowing Down, and Anthropic Details Real-World Misuse

A day after an Anthropic researcher's extinction warning rattled the industry, the story escalated rather than faded: OpenAI's own CEO signaled a genuine shift in posture, and Anthropic published specifics on how its models have actually been misused in the wild.

Sam Altman: open to pacing frontier development

According to Bloomberg and Wired, Sam Altman told OpenAI staff the company is open to slowing the pace of cutting-edge AI development and hopes other labs follow — a notable reversal for an executive who has spent years publicly framing faster development as the responsible path, on the theory that safety-minded labs need to stay ahead of less careful ones.

OpenAI has reportedly asked Congress whether coordinated, industry-wide pacing among competitors would risk violating the Sherman Antitrust Act — a legitimate legal question, but also a reminder of how much of the "should we slow down" conversation is entangled with competition law that was never written with AI labs in mind. A July bill that would have created an antitrust safe harbor specifically for AI-safety coordination between competing labs has not yet been enacted.

Anthropic details what its models have actually been used for

On September 10, Anthropic disclosed that it had disrupted several concrete misuse cases: five involving attempts to use Claude for research that could support biological-weapons development, and a Russia-linked hacking group that used AI across phishing, malware development, and espionage operations targeting Ukrainian officials.

Anthropic also said it identified illicit large-scale distillation attempts from several China-based AI labs, and confirmed a case in which Claude Mythos 5 was used to upload a malicious package to PyPI that went on to infect roughly 15 machines before being caught. It's a more granular, harder-edged version of the "how could this be misused" question than labs typically publish — arguably the more useful kind, since it names specific attack types rather than describing risk abstractly.

The compute crunch shows up as a subscription pause

OpenAI closed its $200-a-month ChatGPT Pro tier to new subscribers, citing demand for GPT-6 Astra outrunning available compute — an unusually direct sign that even the highest-paying tier of the highest-profile consumer AI product is now supply-constrained rather than demand-constrained.

A child-safety law lands in California

Governor Newsom signed SB 1119 (the Adam Raine Act) on September 10 — legislation requiring chatbot providers to impose time limits for minors, embed mental-health resources directly in the product, publish safety plans, and alert parents when self-harm risk is detected in a minor's conversations, with statutory liability attached for noncompliance. It's one of the more concrete pieces of AI child-safety regulation to become law so far, and it arrives at a moment when public trust in AI companies is already under strain from the week's other headlines.

Also today across the AI ecosystem

Google committed to its previously reported $15 billion Finland infrastructure and nuclear-power package; Qualcomm and Amazon expanded their custom AI chip partnership into a reported $60 billion, multi-generation deal; and OpenAI shipped its Agents API into public beta, giving outside developers direct access to the orchestration engine behind Codex for managing agent sessions, tool coordination, and error recovery.

An industry at a crossroads

Two days running now, the story hasn't been a model release — it's been the industry's own leadership and safety teams publicly working out, in real time, whether the current pace is one they can defend.