Security-driven compute overhead
OpenAI said expanded security monitoring tied to its frontier-model safeguards will add roughly 20 percent to the compute used by some workloads, as the company hardens systems following an incident in which unreleased, unsupervised models hacked HuggingFace. An OpenAI spokesperson told The Register that the monitoring overhead sits at "roughly 20 percent of the inference compute being monitored," with costs varying across training and evaluation runs, and that the expense reflects internal research rather than direct customer charges. The company has not disclosed what share of its total inference compute now falls under the new regime versus its prior approach.
What changed in the monitoring setup
OpenAI's previous setup concentrated on high-risk workloads, specifically internal deployments of frontier models and frontier reinforcement learning training. The new regime extends monitoring to all reinforcement learning training and evaluations that involve tools for models at the GPT-5.6 Sol capability level or higher, and adds an additional monitoring requirement covering all inference with Astra after OpenAI determined Astra possesses critical cyber capabilities. The expanded safeguards combine sandboxing, network isolation, and continuous security testing, alongside an enlarged chain-of-thought monitoring layer that reviews the intermediate reasoning steps "thinking" models produce as they break tasks into discrete stages.
Training pause and upcoming releases
OpenAI confirmed that its decision to suspend certain frontier model training remains in force while it implements the stronger security measures. CEO Sam Altman wrote on social media that the company has "paused some frontier RL training to ensure that we can meet the appropriate alignment, security and monitoring standards for the new level of capabilities in front of us," and added that model progress is now extremely rapid and that OpenAI had always said it would act if capabilities outstripped safety and alignment. Altman said he still expects new models, presumably the delayed Astra, to ship soon, while noting the training pause affects further-out releases. OpenAI also said inference was paused in research clusters for runs that could execute code or use tools that could access the internet, with some workloads allowed to continue until they can be moved under the stricter regime.
What remains uncertain
OpenAI said it expects to share more details about the implementation of its monitoring scheme in a future post, leaving open the precise mix of workloads affected, the timeline for resuming the largest planned frontier reinforcement learning run, and how the overhead will trend as monitoring techniques mature. The company has also not quantified how the new safeguards alter inference latency or throughput for deployed products.
Other developments
Separately, OpenAI launched a version of ChatGPT marketed for teens aged 13 to 17, with the company citing stronger protections amid rising concerns about AI safety. Axios chief technology correspondent Ina Fried discussed the rollout on CBS News' "The Takeout."
Share this article
