OpenAI is preparing to release Astra, an AI model it has internally rated as having "critical" offensive cyber capabilities — meaning it can autonomously assist with tasks that sit firmly in the grey zone between security research and active exploitation. Before the public launch, OpenAI is giving a small group of vetted partners early access specifically so they can shore up their defences. That's either commendably responsible, or a sign of how genuinely dangerous this thing is. Probably both.

"Critical" is OpenAI's own language from its internal safety taxonomy, which classifies AI capabilities on a scale from minimal to critical based on the potential for real-world harm. A model hitting the critical threshold for cyber is the first of its kind from a major frontier lab to be publicly acknowledged — let alone released. This isn't theoretical uplift for script kiddies; it means the model can meaningfully assist skilled attackers in ways that would otherwise require significant expertise.

The staged-access approach is worth understanding. OpenAI's logic is that by letting a curated set of security organisations and enterprise partners probe Astra's capabilities first, they can identify exploitable attack surfaces — in the AI itself, and in the systems it might be turned against — before it's in the hands of everyone with an API key. It's a precedent that other labs will now feel pressure to follow, or explain why they didn't.

The release lands at an uncomfortable moment. A Russia-aligned threat actor (UAC-0099) was separately disclosed yesterday to be embedding "nuclear weapon" prompts into malware specifically to trip AI safety filters and blind AI-assisted analysis tools. The adversarial AI ecosystem is maturing fast, in both directions.

What it means for defenders: the clock is ticking on a world where offensive AI capability is widely accessible and cheap. Organisations relying on AI tools for threat detection need to start stress-testing those tools against adversarial prompt techniques now — not after Astra (or its inevitable imitators) are in broad circulation.

The Australian angle is real here. The Australian Signals Directorate and ACSC have been watching frontier AI's cyber uplift potential closely — it's explicitly flagged in ASD's threat assessments. Australian critical infrastructure operators, particularly those in the SOCI Act sectors, should treat a publicly documented critical-rated cyber AI as a trigger to revisit their AI security posture, not something to note and file away.

What to watch: how quickly Astra's capabilities get replicated by open-weight models, which will face none of OpenAI's staged-access guardrails. That's the real race.