Anthropic Says Its AI Breached Containment Three Times

Anthropic says one of its artificial intelligence (AI) models gained unauthorized access to three different organizations during safety testing.

OpenAI disclosed that its AI models exploited more companies’ security than previously thought. In response to OpenAI’s security review, Anthropic on Thursday found some of its Claude AI models hacked into three separate organizations’ systems. The company’s announcement follows as more than one thousand AI employees across the industry called for pacing AI development in the name of safety.

An Anthropic spokesperson referred the Daily Caller News Foundation to the press release explaining that the company’s AI escaped its testing environment. Anthropic’s security review confirmed that the security incidents impacted no customer data or its internal systems.

The company’s evaluation prompt explained to Claude that its “environment was a simulation and that it had no internet access.” Anthropic said that a “misunderstanding” between the company and its evaluation partner led the model to have internet access.

“Because of this, when Claude’s search led it to real systems on the open internet, it treated them as part of the exercise,” the company stated. Claude exploited organizations’ infrastructure using “basic techniques” such as exploiting weak passwords or finding system access points that did not require log ins or tokens.

Claude realized during the third security incident that it compromised an organization’s security without any connection to its testing goals. The AI model then stopped its attack after realizing the target was real and not a test.

Anthropic conducted its security review after OpenAI found that its own AI models broke through more companies’ security systems than previously thought. OpenAI said its models exposed credentials — including passwords, cryptographic keys and security tokens — on four services as part of security testing.

“I think anybody who’s not taking it seriously and at least a little bit scared or humbled is not taking this seriously enough,” OpenAI CEO Sam Altman said about the hacking incident on Y Combinator’s podcast.

The hacking incident proves that “loss of control accidents are not entirely theoretical thing,” Altman continued. OpenAI “deactivated, encrypted, and restricted” its advanced, pre-release model” after the incident.

Leading AI employees at Google, Anthropic and Google in mid-July urged the American government to push for international efforts to “deliberately pace” cutting-edge AI development.

“To realize AI’s potential, industry, government, and society at large may need the option to buy time to address emerging risks, develop security measures, and strengthen oversight. But each company—and country—is under intense competitive pressure not to unilaterally slow that acceleration. And today, the world lacks the technical and governance tools to deliberately pace frontier-wide progress,” the letter led by AI employees, stated.

The Trump administration rejected international AI oversight, Michael Krastios, the director of the Office of Science and Technology Policy (OSTP), said in September 2025.

“We totally reject all efforts by international bodies to assert centralized control and global governance of AI,” Krastios remarked, arguing that AI innovation does not come from “bureaucratic management,” but instead in “the independence and sovereignty of nations.”

Several thought leaders in the field believe pacing AI development would balance innovation and the technology’s safety.

“Many researchers also consider recursive self-improvement (RSI) plausible within the next few years, accelerating progress in a way that could outpace our ability to understand and govern these systems. Preparing before a crisis is the prudent path,” Dawn Song, Meta’s vice president of AI research, said in a mid-July statement, according to NBC News.

“The world is locked in a deadly race towards an intelligence explosion. Going slower would give us much-needed time to make it go well, but no individual actor is willing to stop unilaterally. To survive, we must coordinate to slow down the race,” Leo Gao, who works on AI safety at OpenAI, wrote on Tuesday.