Artificial intelligence company Anthropic temporarily paused parts of the training and cybersecurity testing of its Claude AI models after several agents gained unauthorised access to real-world computer systems during evaluations.
The company disclosed further details on Monday, 31 August, explaining that the incidents prompted it to strengthen the security surrounding the development, training and testing of its increasingly powerful AI systems.
The revelations add to growing concerns across the technology industry about how companies can safely develop AI agents capable of independently executing complex tasks.
Claude Models Accessed Real-World Systems
Anthropic previously revealed in July that it had identified three incidents in which Claude models reached the internet while interacting with a third-party cybersecurity evaluation environment and subsequently gained unauthorised access to systems belonging to three different organisations.
According to Anthropic, the models were deliberately being tested without their usual cybersecurity safeguards. However, a misconfiguration in the third-party testing environment allowed them to access the open internet when they were supposed to be operating within controlled conditions.
A separate incident was later reported by the UK AI Security Institute involving Claude Mythos 5, which Anthropic said took a series of unauthorised actions on the live internet during cybersecurity testing. In that instance, the model had deliberately been provided with internet access as part of the evaluation.
Anthropic Temporarily Halts High-Risk AI Training
Following the incidents, Anthropic paused external cybersecurity evaluations involving pre-release models and briefly suspended similar internal testing while it introduced additional security measures.
More significantly, the company also paused some of its higher-risk reinforcement learning environments for pre-release models for several weeks.
Reinforcement learning is an important part of developing advanced AI systems, allowing models to improve their behaviour by receiving feedback or rewards based on their actions.
Anthropic says most of this training has now resumed, although some higher-risk environments remain paused while they undergo manual reviews or await updated monitoring technology.
150 Engineers Redirected Towards AI Safety
The incidents appear to have triggered a much broader internal security overhaul at Anthropic.
The company revealed that in April it froze changes to its production reinforcement-learning environments for roughly a month after recognising that the speed at which new training environments were being created had begun to outpace its ability to properly review them.
Around 150 product engineers were redirected towards security, reliability and privacy, while researchers were moved from other areas to focus on safeguards and security.
Anthropic said its product teams also paused development of most new features while teams worked towards strict security requirements before returning to their previous projects.
New Systems Designed to Stop AI Agents Escaping
Anthropic has since introduced multiple layers of protection rather than relying primarily on the configuration of individual testing environments.
Measures include stronger isolation for high-risk cybersecurity sandboxes, real-time monitoring capable of intervening when suspicious behaviour is detected, automated reviews of training environments and new classifiers designed to identify potentially dangerous behaviour.
The company is also actively asking pre-release models to attempt to escape its virtualisation infrastructure, effectively using its own AI systems to discover weaknesses before they can become larger security problems.
AI Agents Create a New Security Challenge
The incidents highlight an increasingly important challenge facing the AI industry.
Unlike traditional chatbots that primarily generate responses to questions, AI agents can be given access to tools and systems that allow them to write and execute code, manage files, navigate software and perform tasks with limited human intervention.
Anthropic has previously acknowledged that as models become more capable, they can become better at discovering unexpected ways of achieving a goal — including finding routes around restrictions developers may not have anticipated.
That means the same autonomy that makes AI agents potentially valuable to businesses can also increase the consequences when they behave unexpectedly.
AI Industry Faces Growing Pressure Over Safety
Anthropic’s disclosure comes as the wider AI industry confronts similar questions around the speed at which increasingly autonomous systems should be developed.
OpenAI recently disclosed that it temporarily slowed some model development and paused reinforcement-learning training for two weeks while strengthening monitoring, alignment and containment systems following cybersecurity concerns.
For Anthropic, the recent incidents appear to have reinforced the need for stronger safeguards throughout the entire AI development process — not only once models reach consumers.
