이미지: METAL LAB 생성
Summary
- The White House is reportedly considering expanding its AI oversight framework to require pre-release testing for open models that reach frontier-level capability
- The threshold would be capability on par with Anthropic's Mythos-class models and OpenAI's GPT-5.6, according to a White House official
- A new cooperative structure in which major AI labs would join the oversight program as official partners is also under discussion
- 발표 시점
- 2026년 8월 12일(현지 시각), 백악관 관계자 발언 인용 보도
- 핵심 방침
- 프론티어급 오픈모델을 AI 감독 프레임워크에 포함, 사전 출시(pre-release) 테스트 적용
- 기준 모델
- 앤스로픽 Mythos급 모델, 오픈AI GPT-5.6과 동등한 역량 도달 시 적용
- 추가 검토안
- 주요 AI 랩이 감독 프로그램의 공식 파트너로 참여하는 새 협력 구조
- 우려 배경
- 모델이 자율적으로 국방부(Pentagon)나 글로벌 금융시장을 해킹할 가능성
- 관련 사례
- 오픈AI 차기 모델 아스트라가 내부 평가에서 '치명적' 사이버 역량 임계치 배제 불가 판정(2026년 8월 8일)
No exception for open models
The White House is reworking its AI model oversight system. Reports indicate that an oversight framework previously centered on closed frontier models is now being considered for expansion to include open models that reach frontier-level capability. The news was shared on August 12 by Andrew Curran, who runs the News Amp account, citing a White House official as the original source.
The key issue is the "threshold." According to the White House official, the moment an open model reaches capability comparable to Anthropic's Mythos-class models or OpenAI's GPT-5.6, it would automatically fall under oversight and be required to undergo separate pre-release testing. Until now, regulatory focus has centered on closed models directly deployed and controlled by major labs — but this would extend the same standard to open models, whose weights anyone can download and use.
Why cyber capability is the threshold now
Behind this move lies growing concern over how rapidly AI's cyber capabilities have advanced in recent months. The White House official said the trigger for this action was concern that models could autonomously hack Department of Defense systems or global financial markets.
That concern isn't unfounded. As METAL LAB reported on August 8, OpenAI's next model, Astra, could not be ruled out as crossing the "critical" cyber capability threshold under the Preparedness Framework during internal evaluation. Compared to previous models, including GPT-5.6-Sol, which scored one level lower at "high" in the same evaluation, the pace of capability growth has notably accelerated. OpenAI subsequently applied security controls immediately, including building isolated testing environments and strengthening model weight encryption.
Two days later, on August 10, OpenAI expanded its cybersecurity initiative "Daybreak" and released GPT-5.6-Cyber, a specialized model for authorized vulnerability research. The model reportedly identified previously unknown vulnerabilities in widely used open-source software, including Chrome's V8 engine. With signals piling up on both the defensive and offensive sides that models' cyber capabilities are approaching real-world levels, regulators appear to be moving to extend oversight to open models as well.
What "official partner" status would mean
Separate from expanding the scope of oversight, the White House is also considering a new cooperative structure that would bring major AI labs in as official partners of the program. Until now, government model evaluation has been a regulatory and verification process conducted from outside the labs; if this plan moves forward, labs such as OpenAI and Anthropic could end up inside the oversight framework, helping to design the evaluation standards themselves. However, specific details on the form of participation or which labs would be involved have not yet been disclosed.
Expansion of the oversight framework
| Item | Current | Expansion under review |
|---|---|---|
| Covered models | Closed frontier models | Include frontier-level open models |
| Testing timing | Mainly post-release evaluation | Mandatory pre-release testing |
| Labs' role | Subject to regulation | Official partnership under review |
So what changes
Open models have long faced criticism for being harder to control than closed models, since anyone can download, modify, and redistribute them. If this policy is finalized, developers releasing open models could be required to undergo government pre-verification once their models approach GPT-5.6 or Mythos-level capability. This can be seen both as a brake on the speed and freedom of the open-source AI ecosystem and as an additional safeguard for filtering out model risks before deployment. Still, the specific implementation timeline, the criteria for determining which models qualify, and the scope of participating labs appear not yet to have been finalized.
