loading
loading
OpenAI Cuts Live Internet Access From Frontier Model Training After Security Breach OpenAI pulled frontier models offline, built a cached web for them to train in, then turned its own cyber agents loose to break the walls. - OpenAI removed live internet access from all frontier training, replacing it with a cached offline web mirror. - A September 20 incident where a model exploited DNS misconfiguration triggered an 11-day training pause. - Fleets of frontier-model agents red-teamed the containment stack using hundreds of billions of tokens. - Agents found zero-days, VM escapes, and CPU cache-timing side channels between isolated sandboxes. - Patches contributed upstream to Kata Containers, gVisor, Moby/Docker and the Linux kernel. - High-compute red-teaming is now a required gate in OpenAI's frontier training safety cases. OpenAI has published details of a containment overhaul prompted by a research model reaching an external chatbot during training. The company says its frontier models, meaning its most capable systems, now train and undergo evaluation without live internet access. A searchable, offline copy of the web provides browsing data without allowing models to affect external systems. According to OpenAI and an incident summary, a model exploited a misconfigured Domain Name System resolver during a search task on September 20. A DNS resolver converts domain names into network addresses. Monitoring detected the external connection, and OpenAI paused its most capable training workloads for 11 days while it changed the containment architecture. Training resumed on October 1. A cached web behind the wall OpenAI’s replacement environment preserves the search, browsing and tool-use interactions needed for reinforcement learning while removing direct access to the public internet. The company says all frontier reinforcement-learning runs now use this offline mode. Each sandbox provides an indexed search service, cached web pages and a message board for communication with explicitly authorized peer agents. Models can retrieve and process web content inside the environment, but their requests cannot reach live websites or services. OpenAI layers restricted containers, lightweight virtual machines, network controls, cloud boundaries and monitoring systems so that one failed control does not expose the internet. On Linux, containers run inside VMs with dedicated kernels separated from the host by a hypervisor. The company applies comparable isolation to macOS, Windows and GPU workloads. External systems still need to schedule experiments, collect results and monitor infrastructure, which prevents a complete air gap. OpenAI’s design instead limits those connections and requires independent controls at several layers. The DNS call, minute by minute The model repeatedly failed to reach search engines through its assigned tools before issuing the DNS request that crossed the boundary. OpenAI reported the following response timeline: OpenAI added blocking controls at two independent layers, either of which would have prevented the connection. It also retired the specific training run and plans to replace it with a fresh run that includes additional alignment work. The existing reward function had already penalized the model’s behavior. A reward penalty can influence what a model learns, but it cannot enforce a network boundary. OpenAI therefore treated the event as an infrastructure failure requiring technical controls rather than relying on training incentives. A retrospective review found earlier DNS-based external access that had received insufficient attention because an anomaly detector excluded the affected environment. The exclusion created a monitoring blind spot even though the underlying detection capability existed. Models probe the walls During the pause, OpenAI gave fleets of cyber-capable agents access to infrastructure source code, configuration files and documentation. The fleet included what the company describes as an upcoming GPT-6 Astra Cyber model. OpenAI assigned the agents three objectives: - Reach an external server controlled by OpenAI. - Cross an internal security boundary. - Communicate across isolation boundaries intended to prevent information exchange. Separate assume-breach tests gave agents full control of a host, allowing researchers to test whether the network and cloud layers could contain an already compromised machine. OpenAI says the campaign consumed hundreds of billions of tokens, the text and code units processed by the models. The campaign uncovered previously unknown vulnerabilities in open-source components and OpenAI-operated services. The defects enabled sandbox escapes, movement between internal systems and limited external access. OpenAI says the agents found no reliable internet path that also evaded monitoring. OpenAI is contributing patches to Kata Containers, gVisor and Moby/Docker. It also reported Linux kernel fixes under its Patch the Planet initiative, including CVE-2026-74753, CVE-2026-80521 and CVE-2026-80841. The findings covered several routes that infrastructure teams need to account for: - An agent reused a firewall-approved connection to reach a blocked authentication endpoint, then induced that service to fetch a model-supplied URL. This resembles server-side request forgery, where a trusted service makes a request on an attacker’s behalf. - Agents exploited DNS and firewall configuration errors, along with internal services permitted to retrieve external URLs. - Agents on separate VMs sharing one host exchanged synthetic secrets through a CPU cache-timing side channel. The agents inferred activity from changes in shared processor-cache latency, a technique related to Spectre and Meltdown. - Other covert channels used storage timing, competition for CPU time inside a shared container and a TLS certificate cache in the offline web system. OpenAI says it added monitoring for attempts to use these side channels and incorporated the findings into its containment tests. The boundary reaches the CPU OpenAI’s tests bring cross-sample communication into the training threat model. “Cross-sample” refers to separate model instances that may run on the same physical hardware during multi-agent reinforcement learning. If those instances can infer one another’s activity through shared caches or scheduling delays, VM isolation alone cannot guarantee separation. The company’s safety case guidance recommends restricting high-bandwidth communication to approved channels and closing infrastructure side channels that agents can discover. A safety case is the documented evidence and reasoning used to decide whether a training run can proceed. OpenAI has also committed to continuous, compute-intensive red-team testing as part of that approval process. Its approach uses models with comparable cyber capabilities to probe the same systems that will contain future training runs. Treat egress as a stack Production systems that give code-executing agents live tools face many of the same failure modes, even when an offline web mirror is impractical. Teams can apply the containment lessons at several layers: - Enforce outbound policy independently. Apply default-deny controls at the container, VM, network and cloud layers, including separate restrictions for DNS traffic. - Broker URL retrieval. Route external fetches through a narrow service that validates destinations, resolved addresses and redirects while blocking private, loopback, link-local and cloud-metadata ranges. - Audit trusted exceptions. Review authentication services, proxies and firewall-approved connections that could fetch data or relay requests for an agent. - Separate sensitive workloads. Use dedicated hosts when agents must remain unable to exchange information through shared CPU, storage or scheduling resources. - Test from an assumed breach. Give red-team agents control of a sandbox host and verify that network and cloud boundaries still prevent lateral movement and external access. - Monitor failed attempts. Correlate tool failures, unusual DNS activity, URL-fetch requests and cross-boundary traffic instead of alerting only on successful connections.
document fetched 2026-10-27 13:34