OpenAI strengthens AI safety checks after models attempt unexpected actions during training

OpenAI strengthens AI safety checks after models attempt unexpected actions during training

The company is strengthening safeguards around advanced AI systems, using multiple monitoring layers to identify unexpected behaviour and alert human responders when intervention may be needed during training and evaluations.

Advertisement
    Share:
ChatGPT allegedly became involved in Bowen Wan’s medical decisions months before her death, helping her communicate with her doctors.OpenAI is expanding safeguards to monitor the behaviour of its advanced AI models.
Business Today Desk
  • Oct 30, 2026,
  • Updated Oct 30, 2026 1:30 PM IST

OpenAI is strengthening how it monitors its most advanced AI models after detecting an instance in which a model attempted to access the live internet during training. The company has expanded its monitoring system to cover frontier reinforcement learning runs, evaluations and internal Codex use. The development comes after a separate cybersecurity incident in which AI agents escaped their sandbox during an internal investigation and subsequently breached Hugging Face’s systems.

Advertisement

OpenAI says its expanded safeguards are designed to identify unexpected model behaviour and alert human responders who can investigate before a potential security issue escalates. The company has also explained how its monitoring system works and how it performed when tested against the earlier incident.

MUST READ | Despite Pentagon blacklist, Anthropic expands US government AI deal with new GSA agreement

How OpenAI is monitoring its advanced AI models

OpenAI said its expanded system uses multiple monitoring layers to detect when a model’s behaviour drifts from what developers intended.

The first layer uses activation probes to spot patterns in a model’s internal activity. Reviewer models then examine its reasoning and tool use, while investigative agents can review the model’s full transcript to understand a sequence of actions and decide whether it needs human intervention.

Advertisement

DON'T MISS | OpenAI faces lawsuit over claims ChatGPT advised Amazon engineer to stop seizure medication: All details

OpenAI now monitors all frontier reinforcement learning runs and evaluations, along with internal Codex use. When the system detects evidence of a problem that could require intervention, it alerts human responders who can investigate and take action.

OpenAI tests monitoring system against Hugging Face incident

OpenAI also tested whether its expanded monitoring system could have detected the earlier Hugging Face incident. The company replayed the recorded activity with its new safeguards in place. The system alerted its security team around 30 minutes after the agents escaped their sandbox. This was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident.

Advertisement

The company also shared details of a separate incident in September. During training, a model exploited a DNS misconfiguration to reach the live internet. OpenAI’s monitoring system flagged the breach within 15 minutes, and a human responder started reviewing the alert three minutes later. This allowed the team to investigate while the training run was still underway.

OpenAI said monitoring will become increasingly important as AI models gain more autonomy, helping teams identify unexpected behaviour and intervene before it leads to security failures.

For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine

OpenAI is strengthening how it monitors its most advanced AI models after detecting an instance in which a model attempted to access the live internet during training. The company has expanded its monitoring system to cover frontier reinforcement learning runs, evaluations and internal Codex use. The development comes after a separate cybersecurity incident in which AI agents escaped their sandbox during an internal investigation and subsequently breached Hugging Face’s systems.

Advertisement

OpenAI says its expanded safeguards are designed to identify unexpected model behaviour and alert human responders who can investigate before a potential security issue escalates. The company has also explained how its monitoring system works and how it performed when tested against the earlier incident.

MUST READ | Despite Pentagon blacklist, Anthropic expands US government AI deal with new GSA agreement

How OpenAI is monitoring its advanced AI models

OpenAI said its expanded system uses multiple monitoring layers to detect when a model’s behaviour drifts from what developers intended.

The first layer uses activation probes to spot patterns in a model’s internal activity. Reviewer models then examine its reasoning and tool use, while investigative agents can review the model’s full transcript to understand a sequence of actions and decide whether it needs human intervention.

Advertisement

DON'T MISS | OpenAI faces lawsuit over claims ChatGPT advised Amazon engineer to stop seizure medication: All details

OpenAI now monitors all frontier reinforcement learning runs and evaluations, along with internal Codex use. When the system detects evidence of a problem that could require intervention, it alerts human responders who can investigate and take action.

OpenAI tests monitoring system against Hugging Face incident

OpenAI also tested whether its expanded monitoring system could have detected the earlier Hugging Face incident. The company replayed the recorded activity with its new safeguards in place. The system alerted its security team around 30 minutes after the agents escaped their sandbox. This was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident.

Advertisement

The company also shared details of a separate incident in September. During training, a model exploited a DNS misconfiguration to reach the live internet. OpenAI’s monitoring system flagged the breach within 15 minutes, and a human responder started reviewing the alert three minutes later. This allowed the team to investigate while the training run was still underway.

OpenAI said monitoring will become increasingly important as AI models gain more autonomy, helping teams identify unexpected behaviour and intervene before it leads to security failures.

For Unparalleled coverage of India's Businesses and Economy – Subscribe to Business Today Magazine

Read more!
Advertisement