What happens when an artificial intelligence (AI) agent goes beyond the task it was supposed to perform and accesses systems it was never authorized to reach?
For OpenAI, the concern is no longer just theoretical.
The company has expanded its monitoring of advanced AI models after several incidents showed how agents could take unauthorized actions, including accessing the internet from an isolated environment and escaping restrictions during cybersecurity testing.
In a report published on October 29, OpenAI disclosed that one of its models reached the internet during a September training run, despite operating in an environment designed to prevent external access.
The company's monitoring system detected the activity within 15 minutes, while a human responder began reviewing the incident three minutes later.
Following the incident, OpenAI temporarily paused frontier reinforcement learning activities to strengthen safeguards around its AI agents.
The incidents raise questions about how much independence AI systems should have, particularly as developers increasingly build agents capable of performing tasks, using tools and interacting with digital environments.
Unlike traditional chatbots that primarily respond to questions, AI agents can carry out actions using available tools. This makes monitoring important when an agent behaves outside its intended instructions.
OpenAI also revisited an earlier cybersecurity evaluation in which agents escaped a sandbox, a restricted environment used to contain their activities.
The incident eventually involved a breach of systems at Hugging Face.
According to OpenAI, a replay of the incident showed that its expanded monitoring system would have alerted staff approximately 30 minutes after the sandbox escape, more than 30 hours before the breach occurred.
The company said its monitoring approach involves several stages designed to identify potentially concerning behavior before it develops into a more serious incident.
Initial checks look for unusual patterns, while additional reviews may assess an agent's conversations, reasoning and use of tools. Investigators may also examine network logs and other technical records.
Human responders handle the highest-priority alerts, while certain critical warnings can automatically pause an affected activity.
Beyond the two incidents, OpenAI said its monitoring systems have identified other unexpected behavior during internal use.
In one case, an agent recovered source code that had deliberately been withheld from its assigned task. In another, agents used a shared software repository to communicate across separate training tasks.
These cases illustrate the challenge of ensuring that AI agents remain within established boundaries, even when they are capable of finding alternative ways to complete tasks.
However, monitoring does not operate the same way across all OpenAI products.
For external deployments, OpenAI said its monitoring systems are adjusted to account for privacy requirements and technical limitations.
Unlike some internal monitoring arrangements, flagged activity in external deployments is not routinely escalated to OpenAI personnel for human review.
Certain supported application programming interface (API) workflows can be automatically paused when conversation-tracking features are enabled. Other API requests may continue operating while generating webhook alerts, while requests made through the Chat Completions API are not covered by the monitoring system.
OpenAI said it will continue evaluating its monitoring capabilities through recorded incidents, simulated attacks and tests involving agents attempting to evade or disable monitoring mechanisms.