Unverified claims of ex-DOGE and Cathedral breaches at Anthropic
Article timeline
The articles in this story. Times are ET.
will depueXDepue says AI lab velocity races undermine security against nation-state adversariesAIWill Depue, an OpenAI-linked researcher, argues that CISOs at OpenAI and Anthropic cannot lock down security because leadership prioritizes speed over security in their competition. He says the sprint leaves the results of billions in R&D exposed to national adversaries, calling "security by velocity" workable only for market share, not for the balance of world power. He cites a reported Slack admin-access breach at Anthropic, which he says is underplayed.
- will depueXWill Depue alleges Anthropic underplays a Slack breach by Cathedral
AIWill Depue, an OpenAI employee, says he has heard that Cathedral gained full administrator access to Anthropic's Slack workspace, exposing every message and research secret. He says Anthropic and OpenAI keep much of their internal work in Slack, and compares the incident to OpenAI's 2023 breach through an external app's access token. These are unverified claims from a social media post.
- AI EraNewsAnthropic's internal systems breached by AI, employee salaries exposed
AIThe original title says Anthropic's internal systems were breached by AI and that employee salary data was exposed. The provided text is only two sentences and gives no details on the attack method, the data taken, or the scope of the breach, so this summary cannot confirm those specifics.
- will depueXWill Depue says China has full access to OpenAI and Anthropic's internal systems
AIWill Depue, posting from the account of OpenAI, says China has full access to OpenAI and Anthropic's GitHub, Slack, and docs. He adds that he would not be surprised if Chinese open-source models contain architecture methods stolen from them. He also says a rumor claims Cathedral received a lot of Slack access, possibly all of research.
clem 🤗XEx-DOGE researchers breached Anthropic's systems in under 12 hours using AI agentsAIClement Delangue says concentrating AI behind closed doors is not safe, citing a Washington Post report that ex-DOGE cyber researchers penetrated Anthropic's systems in under 12 hours using AI agents. The breach reportedly reached Slack channels where the company's most sensitive models are discussed.