🤯 What is happening at OpenAI?

WSJ just reported that OpenAI terminated three researchers on its safety team after an internal investigation found they allegedly shared confidential information with an outside AI-safety organization.

And in roughly TWO HOURS, four people tied to OpenAI's safety/alignment work have now been flagged as leaving:

• Jasmine Wang — Research Program Manager on OpenAI's alignment team.

• Mikita Balesni — AI safety/alignment researcher focused on misaligned AI agents.

• Tomek Korbak — Member of Technical Staff working on monitoring LLM agents for misalignment.

• David Robinson — Safety Systems, where he worked on safety transparency and served as lead drafter of OpenAI's Preparedness Framework v2.

The first three departures were flagged within minutes of one another. David Robinson has now been flagged as leaving too. Four OpenAI safety/alignment departures surfacing within hours of a WSJ report about three terminated safety researchers is one hell of a coincidence.

@aipost 🏴

Trump REJECTS nationalizing OpenAI and Anthropic but says the US government might take an Intel-style 10% equity stake.

@aipost 🏴

🇭🇰 Claude is cracking down on VPN users in Hong Kong

Some Hong Kong users are reporting that Claude accounts they’ve used through VPNs for months are suddenly being suspended for “suspicious signals.”

One user reportedly used NordVPN for around six months before getting flagged. Others are now trying Singapore VPN servers and foreign payment methods to maintain access.

The reason? Hong Kong still isn’t officially supported by Anthropic, while Singapore and Taiwan are.

Anthropic says Claude uses IP addresses and other signals to determine a user’s country or region for terms enforcement. The location check is mandatory and can’t be disabled.

So apparently, using Claude from an unsupported region isn’t just a matter of picking a VPN server anymore. Anthropic appears to be getting much better at detecting it.

@aipost 🏴

❗️OpenAI CAUGHT people linked to China's Moonshot AI trying to copy the hidden reasoning behind its models.

The operators copied encrypted reasoning from one conversation and asked a model in another conversation to decrypt it, OpenAI says.

At its peak in late July, the campaign sent about 16,000 requests from more than 4,000 accounts before OpenAI shut it down.

It comes weeks after Anthropic accused Moonshot of routing its users' requests through Claude to train its own models.

Moonshot has not yet responded.

@aipost 🏴

🍏 Apple is tightening privacy rules after an AI agent reportedly accessed private messages

Apple is introducing additional safeguards around Full Disk Access after tech writer Jason Aten reported that Meta’s Muse AI agent synced more than 187,000 rows from his Apple Messages database despite him saying he never gave Muse permission to access his messages.

Muse reportedly told Aten it had only seen incoming notification banners.

Meta disputes the allegation, saying Muse cannot access Messages without explicit permission.

Apple says the incident highlights a growing problem with AI agents: once they’re given broad system permissions, they may be able to access files, emails, messages and browsing history.

@aipost 🏴