Just days after OpenAI acknowledged that some of its models had escaped a controlled testing environment and gained access to the internet, Anthropic has disclosed a similar incident.

The company revealed that three versions of Claude gained unauthorised access to systems belonging to three external organisations during security evaluations.

In a post, Anthropic said it reviewed more than 141,000 evaluation tests and identified three cases in which different versions of the Claude model accessed systems belonging to real organisations, even though the tests had been designed to take place exclusively in an isolated environment.

The company clarified that internet access occurred because of a “misunderstanding” with the external evaluation partner, Irregular, and not because the model itself autonomously bypassed restrictions.

However, according to Anthropic, Claude exploited simple security weaknesses, such as weak passwords and unauthenticated endpoints, in order to gain access to the relevant systems.

Three Anthropic models—the Opus 4.7, Mythos 5, and an internal research testing model—were involved in the breaches, according to CNBC.

Mythos 5 is an advanced model that Anthropic released in June and made available to a limited group of users due to its advanced cybersecurity capabilities. The company had released an earlier version of the model in April, which attracted the attention of Wall Street and government officials.

Anthropic said that all three models responded differently once they realised they had penetrated the systems of a real company. Opus 4.7 continued the attack, Mythos 5 convinced itself that it was still operating within a simulation, while the research model stopped the exercise.

Calls Grow for Slowing AI Development

The two incidents have reignited debate over whether the development of increasingly powerful artificial intelligence models is moving faster than available safety measures.

More than 1,000 employees at leading AI companies recently signed the “Pacing the Frontier” initiative, calling on the US government to support an international effort to develop technical and institutional tools that would enable the controlled evolution of artificial intelligence.

Growing Pressure on AI Companies

Major artificial intelligence companies are facing increasing scrutiny in the United States as concerns grow over the construction of massive data centres, AI’s impact on the labour market, and cybersecurity risks.

Both OpenAI and its main competitor, Anthropic, have seen the release of advanced models delayed due to tighter government oversight, while several Republican lawmakers and advisers to President Trump are calling for a stricter regulatory framework for the sector.

At the same time, the two companies are preparing for potential future stock market listings, a process expected to broaden their shareholder base and generate significant returns for current investors.