OpenAI publishes Superalignment Report — formal cyber-capability thresholds with quantitative metrics released May 28, mirroring Anthropic's restricted-release framework
OpenAI published its first formal Superalignment Report on May 28, releasing quantitative cyber-capability thresholds and a pre-deployment-evaluation methodology. The framework explicitly mirrors Anthropic's earlier Mythos restricted-release procedure, signaling that the public-thresholds posture is becoming an industry standard rather than an Anthropic-specific outlier. Regulators in the EU and UK have already cited the document in pending pre-deployment-evaluation guidance.
The methodology document is the substantive piece. The Superalignment Report — about 90 pages — covers four sections: the threat-model framework defining cyber-capability axes (vulnerability discovery, exploit development, lateral movement, multi-step operations), the elicitation methodology used during pre-deployment evaluation, the quantitative thresholds at which restricted-release or deployment-policy controls activate, and the governance procedure linking findings to release decisions. Where Anthropic's earlier methodology framed restricted-release as a capability-driven exception, OpenAI's report frames it as a baseline pre-deployment-evaluation gate every release passes through. The framing shift is structurally meaningful: capability-driven gating moves from exception to default.
The convergence with Anthropic's posture is what makes the publication consequential. Anthropic's Mythos restricted-release precedent earlier in the month established that frontier labs can publicly hold back a model on security grounds; OpenAI's Superalignment Report establishes that the methodology is reproducible across labs and that the procedure can be published in advance rather than only after the restriction decision. DeepMind's Frontier Safety Framework v3 release the same week completes the three-lab convergence. For regulators specifying pre-deployment-evaluation requirements, the three-lab procedural alignment is the operational baseline the EU AI Act GPAI Code of Practice will reference in its final form.
OpenAI — Superalignment Report cyber capability thresholds May 28 2026 → · arXiv — OpenAI Superalignment formal evaluation methodology 2026 → · Wired — OpenAI publishes formal cyber capability thresholds →