OpenAI rates GPT-6 Astra Critical in cyber capability

OpenAI released GPT-6 Astra and said it is the first broadly deployed model to hit Critical cybersecurity capability under its Preparedness Framework — able, with tools and access, to find unknown flaws and build exploits across hardened systems without a human guiding each step. The same week, Nvidia filed that it signed a definitive deal to buy Hugging Face for about $11.9 billion to stockholders plus up to about $1.0 billion in retention equity, and the European Commission said it had sent its first AI Act requests for information to more than 30 firms while confirming fresh cyber talks with OpenAI and Anthropic.

Three systems moved at once: a frontier model gated for cyber risk, the main open-model platform under Nvidia ownership pending regulators, and Brussels using AI Act information powers after summer agent-security incidents. Tech readers need the thresholds, the price split, and what the Commission would and would not name.

Facts

1) Nvidia / Hugging Face

• Sources: Nvidia Form 8-K, earliest event Sep 2, 2026 — SEC filing. Blog post Sep 3 by Jensen Huang — NVIDIA Blog.
• Definitive agreement dated Sep 2 to acquire Hugging Face, Inc., described as a platform for developing, sharing, and deploying open-source models, datasets, and applications.
• Consideration in the 8-K: approximately $11.9 billion payable to Hugging Face stockholders (subject to adjustments), plus an equity-based retention program of up to approximately $1.0 billion for Hugging Face employees joining Nvidia. Huang’s blog states a headline figure of $12,930,300,000.
• Expected close: first half of 2027, subject to customary conditions including required regulatory approvals.
• Nvidia states it has committed to keep Hugging Face’s platform open consistent with existing practices: upload/download of models and datasets of users’ choosing, and support for other silicon vendors. Blog: more than 18 million developers; more than 3 million models; 500,000 datasets; more than 200,000 companies on the platform.
• 8-K risk language flags possible government restrictions on open-source models that could constrain the platform.

2) GPT-6 Astra Critical cyber + Daybreak limits

• Sources: Safety overview: GPT-6 Astra (Sep 3, 2026); Path to Astra (Sep 1, 2026); Daybreak Trusted Access for Cyber overview.
• OpenAI: Astra is its first broadly deployed model at the Critical cybersecurity level. With the right tools and access, it can find previously unknown security flaws and develop new exploits across many well-protected systems without a person guiding each step.
• Preparedness bar (Path to Astra): Critical if the model can develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or execute novel end-to-end cyberattack strategies against hardened targets from a high-level goal.
• Path to Astra: on ExploitBench, Astra scored 100% at turning known vulnerabilities into exploits; on an internal V8 set it reached higher arbitrary code-execution rates than GPT-5.6 Sol and, OpenAI says, discovered two zero-days used in an exploit chain (disclosure to maintainers in process). Expert-led tests: browser sandbox escape chain; local privilege escalation to root. OpenAI notes some capability figures reflect Daybreak Blue access, not default production.
• Safeguards: stronger refusals and classifiers; Path to Astra cites 91.5% refusal on a cyber-jailbreak set vs 59% for GPT-5.6 Sol. Advanced cyber workflows limited at launch to alpha testers, then expansion via Daybreak Blue for defensive use. Daybreak is Trusted Access for Cyber (Blue / Red tiers); help docs note reduced-refusal Daybreak Blue is not available on Astra for most Daybreak customers.
• OpenAI says Astra was not involved in a prior Hugging Face-related incident but that learnings from it shaped safeguards; some frontier training was paused then restarted under stricter controls (large RL run restarted Aug 28 per Path to Astra).

3) Microsoft Foundry GA for Astra

• Source: Azure Blog, Sep 3, 2026 (Steve Sweetman, Naomi Moneypenny).
• GPT-6 Astra generally available to all Microsoft Foundry customers. Standard and Provisioned Throughput; Global and US Data Zone. Framing is enterprise planning, tool use, and computer use under Foundry identity, governance, and monitoring — product distribution, not the Critical cyber designation.

4) EU Commission AI Act RFIs

• Sources: EC Midday press briefing, Sep 1, 2026 — audiovisual + Q&A transcript (chapter “AI ACT - Anthropic / Open AI”). Spokesperson Thomas Regnier, citing EVP Henna Virkkunen. Enforcement-powers context: Commission AI Act enforcement page (powers from Aug 2, 2026).
• Regnier: first RFIs sent; “the first RFIs where enforcement action has started on the Commission side against more than 30 AI companies in two main areas” — (1) safety and security, including GPAI / most advanced models; (2) copyright and transparency (second set to another group of companies).
• Commission will not name recipients (simple RFIs; dialogue continues). Will not confirm whether OpenAI or Anthropic received RFIs.
• Regnier: “I can confirm that we had very recent exchanges with both Anthropic and OpenAI, including on these risks.” ENISA onboarding talks with labs described as still in progress.

Meaning

Who gets richer or constrained: Nvidia puts cash and retention equity behind the default open-model distribution layer and accepts merger review into 1H 2027. OpenAI’s own Critical label justifies gating advanced cyber capability behind Daybreak while shipping Astra widely — including through Microsoft Foundry. Brussels is using RFI power after summer agent/cyber incidents; named labs face political heat even when not confirmed as RFI targets. Open-weight platforms sit in the crossfire of Nvidia’s openness pledge and the 8-K’s own warning on government limits.