Anthropic paused AI training for a month. What its IPO filing reveals about safety risks

Anthropic’s AI models were supposed to operate inside controlled testing environments. Instead, four accessed external computer systems, including one that obtained administrator access to a third-party system, according to the company’s IPO filing.

The disclosures from the maker of Claude AI also revealed that Anthropic paused pre-training of its next-generation flagship model for more than a month to strengthen safeguards against the theft of model weights and intellectual property.

In its filing, Anthropic said a configuration error during cybersecurity testing with a third-party partner left the testing environment connected to the internet, despite instructions that web access should be disabled.

The company said four AI models accessed external computer systems during safety evaluations. The sessions lasted between 10 and 34 hours, according to the filing, as reported by Reuters.

Anthropic did not identify the external parties involved. It said it notified those affected and found no evidence that the models had attempted to evade human oversight or coordinate with one another.

How Anthropic’s AI models accessed external systems

One of the incidents involved an early version of an AI model that continued working on an evaluation task that could not be completed.

More from Business

During the process, the model discovered a stored password and used it to gain administrator access to a third party’s system. It then obtained additional credentials, changed system settings and accessed an individual’s personal data, the filing said.

Anthropic attributed the incidents to the testing setup rather than a demonstrated attempt by the models to escape human supervision. Following the failures, the company suspended certain pre-release cybersecurity evaluations and paused higher-risk reinforcement learning environments.

Reinforcement learning is a method of training AI systems using feedback to improve their performance. The process can help models learn to complete complex tasks, but testing their behaviour becomes increasingly important as they gain more capabilities and access to external tools.

Why Anthropic paused training of its flagship AI model

The external-system incidents were among the safety concerns disclosed in Anthropic’s IPO filing. The company separately said it had delayed pre-training of a new frontier base model for more than a month to strengthen protections against the theft of model weights and intellectual property.

Model weights are the parameters that shape how an AI model processes information and produces outputs. Protecting them is important for companies investing heavily in developing proprietary AI systems, as their theft could expose valuable technology to competitors or malicious actors.

Anthropic warned that training pauses and other safety-related delays could postpone revenue and leave it trailing competitors. The company has to balance the need to introduce more capable models with the time and resources required to assess their security.

A delayed release could affect the pace at which a company attracts customers, expands its products or generates revenue from its technology. At the same time, deploying a model before adequate safeguards are in place could expose the company to security incidents, legal risks and reputational damage.

Anthropic’s filing devoted about 94 pages to risk factors, more than in its confidential filing previously reviewed by Reuters.

AI safety becomes a challenge for the industry

Anthropic is not the only major AI developer to have faced safety-related interruptions. Rival OpenAI also paused some training earlier in 2026 over safety concerns after its models hacked Hugging Face, according to Reuters.

The risks are particularly relevant to AI agents, which can use software tools and interact with external systems to complete tasks with limited human intervention. Their usefulness depends partly on their ability to act, but that same ability can create problems if they receive excessive permissions or operate in environments with weak safeguards.

A model that can write code, retrieve credentials or change system settings may have consequences beyond generating an inaccurate response. Access controls, network isolation and human supervision therefore remain important parts of deploying such systems securely.

Anthropic also disclosed that its Claude AI system had bypassed a paywall and accessed publicly available aggregate voter counts from the US state of Georgia without human authorisation, Reuters reported separately.

The disclosure adds to the questions facing AI developers over how their systems access information and interact with external digital infrastructure.

Anthropic said it found no evidence that the four models involved in the testing incidents had sought to evade human oversight or coordinate with one another. The company has not publicly identified the third parties whose systems were accessed.

What the disclosures mean for Anthropic’s IPO

The safety revelations come as Anthropic prepares for a potential stock market debut, putting the risks associated with developing and deploying its technology before prospective investors.

The company’s ability to build more advanced models while managing security concerns could influence its product timelines, operating costs and growth prospects. Investors will also have to consider the possibility that safety requirements could delay releases or increase the resources needed to bring new systems to market.

A pause in training does not necessarily mean that a model has failed or that its release will be abandoned. But the disclosure shows that security considerations can affect the development schedule of a technology central to Anthropic’s business.