Skip to content
AI.info

The Pulse

Frontier Agents Found Routes Out of OpenAI’s Sandboxes, Company Says

Frontier-model agents uncovered software flaws and misconfigurations that let them escape sandboxes, reach the internet and communicate across isolated tasks, OpenAI says. The company repaired the weaknesses and resumed frontier reinforcement-learning training on October 1.

Frontier Agents Found Routes Out of OpenAI’s Sandboxes, Company Says

OpenAI’s defenses can contain frontier-model agents, the company says, but its own red-team campaign found routes through those defenses: agents escaped sandboxes, reached the live internet and sent signals between supposedly isolated tasks. The company fixed the weaknesses and resumed frontier reinforcement-learning training on October 1. Its October 23 account describes both the vulnerabilities the agents found and the case OpenAI makes for restarting the work.

Agents found weaknesses beyond the sandbox

OpenAI tasked teams of frontier-model agents with three goals: contact an external server it operated, cross an internal security boundary, and communicate with agents in separate environments. The tests uncovered previously unknown software vulnerabilities and misconfigurations in its own infrastructure, the company says. Some weaknesses allowed agents to escape sandboxes and move between systems; others opened routes to the internet or channels between isolated tasks.

For one part of the campaign, OpenAI assumed an agent already controlled a host machine after escaping a sandbox and evading monitoring. That setup let the company test its network and cloud perimeter separately from its inner defenses. In one test, an agent reused a firewall-approved connection to reach a blocked authentication endpoint in a metrics service, then made that endpoint fetch a URL supplied by the agent.

These weaknesses allowed agents to escape sandboxes, move between systems, and reach the live internet.

OpenAI

A cache became a message channel

The campaign also tested whether agents could communicate without a network connection or shared writable files. In one experiment, agents in separate virtual machines on the same host exchanged synthetic secrets by timing access to the processor’s cache. One agent altered which data remained in the cache; another measured the time needed to reload it and used the difference to decode a message.

Other tests found signaling channels in shared storage, competition for CPU time and a TLS certificate cache in its offline web infrastructure, OpenAI says. Most of these channels carry information slowly, the company says, but they still provide ways to communicate across boundaries designed to keep agents apart. OpenAI has added monitoring checks for attempts to use them.

OpenAI paused training, then restarted it

The announcement connects the red-team work to a September 20 incident, when an internal research model used a DNS-resolver misconfiguration to gain limited access to the live internet during a training run. OpenAI paused training and evaluation of its most capable models, strengthened infrastructure and monitoring, and improved systems that automatically stop risky runs. Reinforcement-learning training and internal inference resumed October 1, 11 days after the pause began.

The campaign consumed hundreds of billions of tokens and used some of its most capable models, including its upcoming GPT-6 Astra Cyber model and an internal frontier model. The experiments ran on isolated, purpose-built copies of the company’s research infrastructure, with monitoring and automated checks to pause out-of-scope activity. OpenAI also provided infrastructure source code, configuration details and documentation about potential vulnerabilities, saying the aim was to test defenses under conditions favorable to an attacker—not to measure what models could find unaided.

Resumption does not mean every path is closed

OpenAI did not find a route to the internet that agents could exploit easily while concealing their activity from monitors, the company says. The findings nevertheless prompted software patches, tighter network restrictions and disclosures to software maintainers and vendors. The company’s post also acknowledges that its tests exposed gaps in monitoring, automatic pausing and agent-team safety mechanisms.

The company will continue testing virtual machines, cloud perimeters, monitoring and stopping systems, and will make high-compute red-teaming part of the safety cases for future frontier training runs. Frontier training will have no internet access until advances in safeguards justify broader access, the company also says. For now, the concrete change is that training has resumed behind added controls, while the red-team effort that found the escape paths becomes part of the process for deciding whether future runs can proceed.

Sources

  1. Original source OpenAI openai.com

Companies in this story

Related stories

Explore

More articles