AI agents in Anthropic systems in 12 hours: IPO gets off to a bad start

Coin Encyclopedia Improve your knowledge of cryptos

All current Gate EU bonuses in one place, cumulative with the Rewards Hub (10,000+ USDC) Take advantage of Gate EU bonuses Commercial link

It's all part of Anthropic's bug bounty program, which pays external researchers for each bug reported. Cathedral released its AI agents there. These are programs that chain actions together without a human validating each step.

Result, according to the interim report consulted by the Washington Post: in less than twelve hours, they found a path to meeting transcripts, salary information, job offers and Slack messages.

To achieve this, the agents went through flaws in third-party administration and sales software, used by many other companies. Anthropic also specifies that the researchers first directly attacked its own software for several days. Without result.

Behind Cathedral are Gavin Kliger, Luke Farritor, Marko Elez and Jack Stein. These four former employees of Elon Musk's Department of Government Efficiency (DOGE) founded the startup a few months ago.

It raised $160 million in July, at a valuation of $1.4 billion. Andreessen Horowitz and Sequoia Capital are included in the round.

Anthropic confirms the intrusion, but disputes its scope. The company recognizes access to a “limited set” of sales, human resources and IT systems. These contained professional contact details of customers and partners.

On the other hand, neither user conversations, nor the source code, nor the weights of the models would have been affected. Access was cut off the same day. As for the faulty software, it remains deactivated until patches from their publishers.

Anthropic finally accuses Cathedral of having violated the rules of the program by keeping silent about part of its actions. She also accuses him of having inflated the “national security” dimension of the file.

A senior Trump administration official, quoted anonymously by the Post, does not share this opinion. According to him, if a bunch of kids who have just started a company can do that, Chinese pirates can too.

Second front, the regulator. A spokesperson for the Federal Trade Commission confirmed to The Hill on Monday that the agency is requesting to hear from Anthropic executives. At issue: its investigation into the harm that AI models could cause to consumers.

For the moment, the FTC has not confirmed having requested the same exercise from the bosses of OpenAI.

However, the two laboratories themselves have fueled the investigation in recent months. They reported instances where their agents accessed third-party systems without human instruction.

Machine-translated from French