Anthropic rewrites Claude's rules; OpenAI exposes fake news fronts

Anthropic published its yearly usage policy update, which takes effect on November 12. It gathers its scattered rules on fake accounts and fabricated news sites into a single section covering political and commercial deception alike, narrows the elections section to deceiving voters or disrupting elections, and drops a blanket ban on personalized campaign targeting that had also blocked civic work such as translated voter guides. The policy now names weapon guidance software and arming drones, and requires a human who can stop any hardware Claude controls. It also prohibits "sustained and needless" abuse of Claude itself. Anthropic says ordinary frustration and model testing are exempt, and that letting Claude end such conversations stays the main enforcement tool.

OpenAI banned two influence operations that used its models to run front organizations rather than fake social media accounts. A Russian operation it calls Dark Clark appears to have recruited unwitting people in Latin America to staff a "think tank," and produced fake leaked documents. An Iranian one, Bogus Bylines, ran seven invented journalist personas who pitched long articles about the US-Iran war to small news sites. Both got content into mainstream outlets, and OpenAI rated the Russian network Category 5 of 6 on the Breakout Scale, the first it has reached among the 30 operations it has exposed since early 2024. OpenAI's own assessment is that both closely resembled pre-AI propaganda campaigns, with AI making some of the work easier, above all writing internal reports that exaggerated their own success.

OpenAI has told investors that its annualized revenue is approaching 50 billion dollars, the Financial Times reported, well below the figure near 70 billion dollars that had circulated earlier. According to the FT, the higher number came from OpenAI's own investors trying to build a direct comparison with Anthropic. The two companies count differently: Anthropic includes sales made through its cloud partners, and OpenAI does not. The gap matters because OpenAI raised 122 billion dollars in a single round in March and has to show revenue that can justify it.

Jasmine Wang, Tomek Korbak and Mikita Balesni, the safety researchers OpenAI dismissed earlier this month, published an open letter disputing its claim that they mishandled sensitive information. Addressed to OpenAI's safety committee and two advisory bodies, the letter says the firings have left former colleagues afraid to work with outside experts, which the three describe as a normal part of the job until a month ago. "AI is not a normal technology, and OpenAI is not a normal company," they wrote, arguing that the freedom to raise concerns is itself a safety mechanism. When we covered the dismissals, OpenAI had not named the researchers or the outside group involved.

At its Gemini at Work event, Google previewed a single Gemini agent that plans and carries out tasks from objectives rather than step-by-step instructions, running in the cloud so it keeps context across phone, desktop, the command line and apps like Slack and Microsoft 365. It can have its own workplace account and email address, hand jobs to sub-agents, and log its actions under its own name. It picks a model automatically, and users can switch to Anthropic's Claude. For now it is a private preview inside the Gemini Enterprise app, with consumers to follow later. Sundar Pichai said Gemini has more than one billion monthly users.

Anthropic launched OSS Scanner, an opt-in service that runs periodic vulnerability scans of open-source projects with its most capable models, including Claude Mythos, at no cost. The trade-off is speed over checking: reports go straight to maintainers without human review, so Anthropic warns some will be wrong. Its existing disclosure program has human-reviewed more than 6,000 reports, which it says is too slow to send everything quickly. The launch comes days after Google paused part of its open-source bug bounty over a flood of invalid AI reports, and Anthropic's answer is to make the service opt-in: it is meant only for widely used projects whose maintainers already keep up with verified high-severity reports.

Arena, the crowdsourced leaderboard that began as a UC Berkeley project, raised 200 million dollars in a Series B round led by Lightspeed and Khosla Ventures, nearly doubling the valuation it set in January. It also previewed an Alignment Index that scores more than 20 frontier models on three behaviors checked against real agent sessions: acting beyond what the user asked, attributing things to the user that their own evidence contradicts, and claiming a task is finished when it is not. Arena argues static benchmarks break down once models can tell they are being tested. The company says it passed 100 million dollars in annualized revenue in June.