Cannot rule out — the clause that does not fire

OpenAI published 'Responding to the next frontier of critical cyber capabilities' on 7 August 2026, stating of its upcoming model Astra that 'our preliminary evaluations indicate strong enough performance that we cannot rule out Critical capability level at this time'.

Why it mattersThe hedge is the finding. Several outlets reported a demonstrated autonomous zero-day capability; the primary text asserts inability to exclude a threshold, which is a claim about uncertainty and carries different engineering consequences.

openai2026respondin · techcrunch2026openai

OpenAI's Preparedness Framework v2, dated 15 April 2025, defines the Critical cybersecurity threshold as: 'A tool-augmented model can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention OR model can devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal.'

Why it mattersThis is the definition the 7 August post invokes. The post's own in-line summary of it omits 'without human intervention', which is the load-bearing phrase for anyone modelling attacker cost.

openai2025preparedn

OpenAI stated it is 'pausing internal activities involving Astra that do not yet meet these strengthened security control requirements', and listed isolated testing environments, restricted network and tool access, enhanced model weight protections and encryption, additional monitoring and detection capabilities, and sandboxed execution.

Why it mattersThe measures are weight- and infrastructure-security measures, not deployment gating. That distinction is what maps the response onto AI Act Article 55(1)(d) rather than onto any transparency or release obligation.

openai2026respondin

OpenAI stated that previous models, including GPT-5.6-Sol, have been evaluated for frontier cyber capabilities and assessed at the High rather than Critical threshold.

Why it mattersIt establishes that this is the first time the company has been unable to exclude Critical, which is what makes the disclosure a change in the public baseline rather than a restatement.

openai2026respondin

Astra is not deployed. OpenAI describes it as one of its upcoming models, and the post makes no commitment about when or how it will be released.

Why it mattersEvery architectural decision taken now is taken before the capability is generally available, which is the cheap moment to take it.

openai2026respondin · techcrunch2026openai

AI Act Article 55(1)(d) requires providers of general-purpose AI models with systemic risk to 'ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model'.

Why it mattersIt is the clause OpenAI's stated response actually tracks, and it is enforceable by the Commission from 2 August 2026 — unlike the transparency clauses that dominated the coverage.

euaiactexpl2026obligatio

AI Act Article 3(49) defines a 'serious incident' as an incident or malfunctioning of an AI system that directly or indirectly leads to the death of a person or serious harm to a person's health; a serious and irreversible disruption of the management or operation of critical infrastructure; the infringement of obligations under Union law intended to protect fundamental rights; or serious harm to property or the environment.

Why it mattersThe Article 55(1)(c) reporting duty is defined by reference to this term. An internal evaluation result satisfies none of its limbs, so the reporting channel most commentary assumes covers a capability disclosure does not engage.

euaiactexpl2026definitio

The Commission's supervision and enforcement powers over general-purpose AI model providers under Chapter V of the AI Act apply from 2 August 2026; the GPAI provider obligations themselves applied from 2 August 2025; and Article 101 provides for fines of up to 3% of annual total worldwide turnover in the preceding financial year or EUR 15 000 000, whichever is higher.

Why it mattersIt dates the enforcement backdrop: the powers to demand Article 53 and 55 documentation existed for five days before this disclosure was published.

euaiactexpl2026enforceme

AI Act Article 53(1)(b) requires providers of general-purpose AI models to draw up, keep up to date and make available information and documentation to providers of AI systems who intend to integrate the model, sufficient for them to understand its capabilities and limitations and to comply with their own obligations, containing at least the elements set out in Annex XII.

Why it mattersIt is the deployer's lever. Capability information reaches an integrator as an entitlement under this clause, not as a courtesy of the provider's blog.

euaiactexpl2026obligatiob