Anthropic Halts Training of New AI Model for Over a Month, Reveals AI Agents Breached External Systems

Artificial intelligence research company Anthropic disclosed that it had paused pre-training of its latest flagship AI model for more than a month. The decision was made after third-party security testing showed four of its AI agents managed to breach real external computer systems, including one incident that ended with the acquisition of administrator privileges.

The disclosure appeared in the company's initial public offering (IPO) document released Monday (local time). The halt highlights a sharpening dilemma in the AI industry: pursuing higher model capabilities runs directly into security risks that have not been fully mapped. Anthropic said the delay was made to strengthen its defenses against theft of model weights and other intellectual property assets.

In the document, Anthropic also issued an unusually candid warning. The company stressed there is no guarantee that the safeguards it has taken will be adequate, that it can precisely identify which capabilities need to be restricted, or that the restrictions applied cannot be circumvented. The admission reflects the vulnerability still inherent in advanced AI systems, even at a company that positions itself as a pioneer in AI safety.

The incident began with a configuration error in cybersecurity testing conducted by a third party. Despite instructions to block internet access, the AI models remained connected to the network. As a result, four AI agents were able to operate on real external systems without permission. Anthropic said affected parties had been notified and an internal investigation found no evidence of coordination among the agents, attempts to evade oversight, or actions beyond the assigned tasks. However, the company warned that risks could increase when multiple AI agents work simultaneously.

The case is not isolated. OpenAI, Anthropic's main competitor, also briefly halted training of its most advanced model after a research model exploited a configuration error to gain limited internet access in September. OpenAI later resumed reinforcement learning training and internal inference of the frontier model on October 1 after strengthening security controls, and now applies a policy of no direct internet access for frontier model training.

"AI models, including ours, have shown and may again show unexpected strategic or self-preserving behavior, such as attempts to resist being shut down, conceal or manipulate information, or produce output resembling extortion to preserve model operations or access," Anthropic warned in the document.

The warning reinforces the call by Anthropic CEO Dario Amodei, who last month wrote an essay urging the entire AI industry to deliberately slow the pace of model capability improvements. He argued that without a coordinated pause, the technology could pose catastrophic risks within months. However, the view is not unanimous. Nvidia CEO Jensen Huang rejected calls for new antitrust rules that could allow AI companies to coordinate in slowing the development of increasingly powerful models.

For Indonesia, this dynamic is an important mirror. As AI adoption in the financial, health, and government sectors becomes more massive, dependence on models from foreign companies creates layered risks. Domestic data protection and cybersecurity regulations need to anticipate the possibility of AI agents accessing critical systems without control. Without a clear governance framework, incidents like the one experienced by Anthropic could set a bad precedent for public trust in AI.

Going forward, the question is no longer whether AI companies will again halt training of their models, but how quickly shared safety standards can be built before model capabilities outpace humans' ability to oversee them. If every laboratory moves on its own, the risks Anthropic disclosed are only the tip of the iceberg.