OpenAI says preliminary evaluations mean it cannot rule out Astra having Critical cyber capabilities under its Preparedness Framework—not that this threshold has been confirmed.
Why it matters: The announcement links a preliminary capability warning to concrete restrictions on development and testing. For readers assessing the claim, the key distinction is that Critical capability has not been confirmed, and the capture does not establish how effective the announced safeguards are.
HarnessSafe traces how attacker-influenced content can persist across agent systems and affect a later benign request.
Why it matters: Safety evaluations should examine carrier-specific traces and harness-model configurations, not only final attack-success rates: those rates can miss how far a risk propagated and where containment occurred.