Trump Administration Gates Frontier AI Releases via Cybersecurity Vetting

OpenAI and Anthropic both deferred model availability to administration-approved partners following an executive order authorizing up to 30-day national security reviews of advanced systems. OpenAI framed the GPT-5.6 Sol limits as temporary safeguards around vulnerability discovery capabilities, while Anthropic restored partial Mythos 5 access only after two weeks of offline compliance. Procurement and directive records show the reviews now directly control which entities receive early capability exposure.

Contract awards and prior Anthropic disclosures indicate Mythos 5 demonstrated automated flaw-finding performance that exceeded prior thresholds for dual-use cyber risk. David Sacks' public statements confirm the April briefing triggered cortisol spikes inside the administration, converting a voluntary framework into de facto gatekeeping. No independent CVSS-style scoring or exploit telemetry has been released to justify the differential treatment between the two labs.

The pattern marks a structural shift: commercial rollout timelines now hinge on geopolitical customer approval rather than technical benchmarks alone. This creates an uneven playing field where San Francisco labs must navigate Washington appointee discretion without statutory process or congressional oversight. Smaller competitors lacking similar relationships face de facto exclusion.

Next reporting period will show whether the 30-day window expands into recurring pre-release holds and whether approved partner lists correlate with specific defense contractor procurement vehicles.