Fired OpenAI researchers deny any leak⁠.

OpenAI's security team asked three of its safety researchers to open up their personal phones during a leak investigation, then fired all three a week later, The Wall Street Journal has reported². The three, Tomek Korbak, Mikita Balesni and Jasmine Wang, deny sharing confidential information outside the company¹. They say OpenAI never told them in writing which information they were supposed to have mishandled¹.

On the phones, security officers saw that some of the researchers had turned on disappearing messages in their Signal chats with Ryan Greenblatt, an outside safety researcher². Investigators named Greenblatt as the suspected leaker, and he denies it². OpenAI said Signal was not an approved tool, but that it was not the only reason for the firings². Korbak called disappearing messages a "basic security best practice" in Silicon Valley².

OpenAI told the Journal that an internal investigation found the three had broken its policies on sensitive information². The company said the firings were "not for raising safety concerns or speaking out"². Wang said she was told she was fired for repeatedly opening an executive's email, access she says she had asked OpenAI to remove nearly a year earlier². She said she opened one sensitive email by mistake and reported it at once².

The inquiry followed a 1 September report in The Information that Astra, OpenAI's latest model, was more powerful but showed less of its thinking². The three had spent the summer pushing to keep that written reasoning, the chain of thought, open to scrutiny². In a letter to OpenAI's board, they urged the company to work more closely with outside safety organisations¹.

Korbak says colleagues are now afraid to talk to outsiders or to raise safety concerns inside the company¹. The researchers also worry the case could put independent safety auditors off working with AI companies altogether¹.

Who gets to check the checkers

We cannot tell you who is right, and nobody outside the room can either. OpenAI says its rules were broken, the three say they were never told what they leaked, and the man named as the source says the accusation is false. None of the reporting carries a British angle, and no UK body has said anything about it.

Two things are worth taking from it. If you buy AI from one of the big labs, part of what you are trusting is that outsiders can look at how its models behave, and this is a fight about exactly those people. Closer to home, a rule nobody wrote down, about which chat apps staff may use and what they may share with outside partners, is the kind of gap that turns an ordinary conversation into a sacking. Writing it down now is cheaper than arguing about it later.

Also today

• OpenAI lowers the valuation it wants to $1.25tn

  OpenAI has cut the valuation it is seeking in a new round of up to $30bn to $1.25tn, from $1.4tn, after some investors pushed back on the higher price³.

• China clears DeepSeek and MiniMax to take models abroad

  A state media account says DeepSeek and MiniMax have registered under China's proposed AI export controls and won rights to release models overseas, while most AI technology stays free to export⁴.

• Anthropic says one agent often beats a team of subagents

  Anthropic says a single agent that loads context as it needs it often beats setups of many specialised subagents on quality, cost and speed, and is offering a kit to help teams switch⁵.

• US Treasury proposes wider curbs on investment in Chinese AI

  The US Treasury has proposed rules that would widen limits on American investment in chips, AI, quantum computing and biotechnology in China, Russia and other named countries⁶.

• Nvidia releases a cheap labeller for robot training video

  Nvidia has released Flora, an open tool it says labels an hour of robot training video for about 72 cents, against $88 with OpenAI's GPT-6 Astra⁷.