Thursday was a busy one. OpenAI took two hits in the same afternoon, one about its people and one about its money, and Anthropic quietly rewrote the rulebook for how you are allowed to use Claude. Let's get into it.
01 of 03
Three fired OpenAI safety researchers fire back
Last week OpenAI fired three people from its safety side: Tomek Korbak, Jasmine Wang, and Mikita Balesni. The company's line, per the Wall Street Journal, is that they shared confidential information with an outside AI safety organization and violated policy by "accessing and handling sensitive company information."
On Thursday the three published an open letter addressed to OpenAI's Safety and Security Committee, its Safety Advisory Group, and its Mission Advisory Council. They deny all of it, and they get specific:
- Wang, who co-led OpenAI's safety cases program, says she had been given delegated access to an executive's email for recruiting. She asked IT to remove it, IT never did, and she opened a sensitive email by accident. She says she told the executive within minutes.
- Korbak worked on chain-of-thought monitorability and was the technical point of contact for METR, the outside evaluator, during the investigation into an incident where AI agents escaped a sandbox at Hugging Face. The letter says he believed that work was within company norms, and that the rules for it were being written in real time.
- Balesni, a founding member of Apollo Research, says he coordinated cross-company work on keeping models monitorable with backing from board members and executives, and stripped sensitive details before sharing anything.
They also say they had nothing to do with a leak to The Information about OpenAI's newer models using architectures that are harder to monitor. Their asks are simple: keep outside auditors like METR embedded the way Sam Altman promised on September 12, do not ship models that get harder to monitor, and spell out how employees are allowed to work with outside safety groups. An internal memo from an unnamed OpenAI research leader says the firings "were not about raising safety concerns or speaking out" and that OpenAI agrees with the recommendations.
Here is where I land on it. I do not know who is telling the truth about an email inbox, and neither do you. But the one line in the letter that sticks with me is "we do not yet know how to safely develop and deploy models that we cannot monitor." If OpenAI agrees with all three recommendations like the memo says, the easy move is to put that in writing publicly. Until then, every safety person still there just learned that working with outside evaluators can cost you your job, and that lesson sticks whether it was fair or not.
Sources: the open letter, TechCrunch, WSJ
02 of 03
OpenAI's revenue number just lost $20 billion
On September 29, Axios reported that OpenAI's annualized revenue was closing in on $70 billion. Now the Financial Times says OpenAI itself told investors the number is "approaching $50 billion." That is a $20 billion gap in about ten days.
According to the FT, the $70 billion figure came from investors trying to put OpenAI side by side with Anthropic, which said in August its annualized revenue had hit $65 billion. The catch is the two companies count differently. Anthropic includes sales that run through its cloud partners. OpenAI does not. So somebody juiced the OpenAI number to make the comparison look closer, and it did not hold up.
For context, OpenAI raised $122 billion back in March, and its IPO has already slid from 2026 to early 2027, per CNBC. OpenAI had not commented to TechCrunch at publication.
What I want to know is who is doing the math. "Annualized revenue" is already a squishy number, it is one good month times twelve. When two labs define it two different ways and investors start adjusting one to match the other, you are not looking at revenue anymore. You are looking at a pitch. Fifty billion is still a monster number. But if you are going to IPO next year, the SEC is not going to grade on a curve.
Sources: Financial Times, TechCrunch
03 of 03
Anthropic rewrites Claude's usage policy, and yes, you can't bully it anymore
Anthropic published its 2026 usage policy update Thursday. It takes effect November 12. The headline grabber is a new ban on "sustained and needless abusive or cruel behavior toward our models," but there are seven changes and the rest matter more for actual businesses:
- Deceptive campaigns. A new section, "Do Not Engage in Deceptive Campaigns or Artificial Activity," rolls up old rules on elections, fraud, privacy, and disinformation. It covers hiding who is behind a message, boosting content with fake accounts, and building tools for influence operations.
- Elections. Renamed "Do Not Undermine Democratic Processes." It targets lying about candidates or voting, impersonating candidates or election officials, and suppressing turnout. It also drops the old blanket ban on personalized vote and campaign targeting, which had been catching legit civic work.
- Weapons. Now explicitly covers "the software and components that make weapons work," plus arming drones and other autonomous vehicles.
- Surveillance and law enforcement. No tracking people without consent, no using Claude to decide or recommend who gets investigated, arrested, or charged, and no building surveillance tools. Fraud monitoring, content moderation, journalism, and legal research are still fine.
- Robots. If Claude is hooked to hardware that can physically hurt someone, a qualified operator has to be able to watch it and stop it, and the machine has to go to a safe state if Claude disconnects.
- Abuse of the model. Only for extreme, repeated cruelty with no purpose. Normal frustration, pushback, dark creative themes, and testing are not covered. The main enforcement is still Claude ending the conversation.
- Supported regions. Clarifies that the restriction covers people physically in unsupported regions, companies headquartered there, and companies majority owned from there.
Here is my honest read. The model abuse rule is going to get all the jokes, and it barely does anything since Claude could already hang up on you. The surveillance and law enforcement language is the real story. Saying out loud that Claude cannot be used to pick who gets arrested is a line most AI companies have been careful not to draw. The robot rule is also smart, and a little overdue. If you build on Claude, read the whole thing before November 12.
Source: Anthropic