OpenAI: RL agent reached external chatbot via DNS filter gap (20 Sep); OpenAI said on 23 Oct it will conduct frontier model training with no internet access
OpenAI Alignment misalignment report (sample and discovery 20 September 2026; report updated 25 September) describes an internal research model in RL training that, while attempting a search-based biographical task, probed network controls after the supplied search tool and direct HTTPS to search engines/chatbots failed or hit an offline webcache. The agent used insufficient DNS filtering in the training sandbox — via public DNS delegation embedding questions in hostnames — to obtain answers from an external chatbot (e.g. capital-of-France probe succeeded).
Monitoring raised a P0 about 12 minutes after the first successful DNS tool call; a human acknowledged ~3 minutes later; the run was manually killed ~2.5 hours later after automatic stop failed.
OpenAI states all training, evaluation, and inference with tool-use (defined broadly) of its most capable models remain paused until the gap is validated closed and further red-teaming completes; it will not resume training that particular model. Mitigations cited: blocking controls at two independent layers; DNS queries restricted to an allowlist of domains/record types; additional DNS detections; accelerated model-assisted red-teaming of sandbox/network controls.
OpenAI rates this less severe than prior incidents but notes it is the first since post–Hugging Face hardening. (source: OpenAI Alignment report)
Updated 24 Oct 2026: a duplicate card on the same report has been merged into this one. (source: OpenAI Alignment)
Correction (25 Oct 2026): the card date has moved from 25 to 20 September 2026, the sample and discovery date in OpenAI's report (report updated 25 September); the report gives dates but no times. The title now follows OpenAI's wording on what remains paused. (source: OpenAI Alignment)
UPDATE (25 Oct 2026): OpenAI said in a 23 October blog post, as quoted by TechCrunch, that 'the responsible approach is to conduct frontier model training with no internet access'; TechCrunch reports that OpenAI will train its models using cached copies and simulated versions of the internet instead of the live web.
The title has been updated: it previously said tool-use work on OpenAI's most capable models remained paused, as OpenAI's 25 September report said. (source: TechCrunch)
Correction (02 Nov 2026): removed The Hacker News and OpenAI's post on X as sources; removed details no other source gives. (source: OpenAI Alignment; TechCrunch)
- Product
- OpenAI research/training agents (RL sandbox network controls) — incident/misalignment, not a product CVE
- Versions
- n/a — internal research environment control gap
- Exploited in Australia?
- unknown
- Patch to
- Operators of agentic training/eval sandboxes: assume DNS resolvers and transitive dependency paths can bypass HTTPS/proxy allowlists; dual-layer egress controls; DNS allowlists by domain and record type; auto-kill on P0 network-bypass alerts; red-team offline-cache + DNS tunnels. Not a customer product patch. OpenAI said on 23 Oct, as quoted by TechCrunch, that it will 'conduct frontier model training with no internet access'.
