News follow-up · Frontier models
Frontier model safety: what happened since
In May 2024 one company's board committee was the news. Since then the oversight of frontier models has gained a government layer: pre-release testing agreements in 2024, a renamed federal centre, and in June 2026 an executive order that asks developers to hand over their most capable models for up to 30 days before release, voluntarily.
What this site reported
On 28 May 2024 the earlier version of this site reported that OpenAI had formed a board safety and security committee as it began training the successor to GPT-4.
What the record shows now
Government testing began in 2024. On 29 August 2024 the U.S. AI Safety Institute at NIST announced memoranda of understanding with Anthropic and OpenAI under which it would "receive access to major new models from each company prior to and following their public release", with feedback on safety improvements given in collaboration with the U.K. AI Safety Institute.[1]
The company committee became one layer of three. OpenAI's Preparedness Framework, version 2 of 15 April 2025, describes the chain of decision: an internal Safety Advisory Group reviews the evaluations and recommends safeguards, company leadership approves or rejects the recommendation, and "our Board's Safety and Security Committee provides oversight of these decisions".[2] The framework tracks three categories of capability, biological and chemical, cybersecurity, and AI self-improvement, and commits the company not to deploy a model that reaches a High threshold until the associated risks "are sufficiently minimized".[2]
The federal centre changed name and remit. NIST's programme page now presents the Center for Advancing Innovation and Standards for Super Intelligence, CAISSI, as "industry's primary point of contact within the U.S. government" for testing and research on commercial systems.[3] Its stated work includes voluntary agreements with developers, unclassified evaluations focused on "demonstrable risks, such as cybersecurity, biosecurity, and chemical weapons", and assessments of adversary systems for backdoors and covert behaviour.[3] The news items on the page are assessments of individual foreign models, one of them joint with the U.K. institute on 23 July 2026, and a 16 October 2026 note on incident readiness in training and evaluation.[3]
An executive order on pre-release access. Executive Order 14409 of 2 June 2026 directs the Treasury, the Department of War through the National Security Agency, and Homeland Security through CISA, with NIST consulted, to build within 60 days a classified benchmarking process for the cyber capabilities of AI models and to define a "covered frontier model".[4] It asks them to design "a voluntary framework" under which developers could give the government access to such models "for a period of up to 30 days before they plan to release such models to other trusted partners".[4] Section 3(c) says nothing in the order authorises "a mandatory governmental licensing, preclearance, or permitting requirement" for releasing new models.[4]
What it means for public bodies
The question of 2024 was who checks a frontier model before release. The answer on the record in October 2026 is: the developer's own group and board committee, under a published framework;[2] the federal centre, under voluntary agreements;[1][3] and, if the June 2026 framework is built, a classified cyber benchmark with up to 30 days of pre-release access, still voluntary.[4] None of the four documents gives a count of models tested or a result that an agency could cite when buying one.
See also: Frontier model risk ratings: what happened since, AI self-regulation: what happened since and Generative AI in government.
Sources
- National Institute of Standards and Technology, news release, U.S. AI Safety Institute Signs Agreements Regarding AI Safety Research, Testing and Evaluation With Anthropic and OpenAI, 29 August 2024 (page updated 4 May 2026). Read 29 October 2026.
- OpenAI, Preparedness Framework, version 2, 15 April 2025. Read 29 October 2026.
- National Institute of Standards and Technology, Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), programme page with news items to 16 October 2026. Read 29 October 2026.
- Federal Register (govinfo), Executive Order 14409, Promoting Advanced Artificial Intelligence Innovation and Security, signed 2 June 2026, published 5 June 2026. Read 29 October 2026.
How this page was checked: the four documents were read in full on 29 October 2026. OpenAI's own announcement pages were not consulted; the record used for the committee's later history is the framework document on the company's file server.