Overview
U.S. executive-branch effort (June 2026 EO → Aug 2026 industry handoff) to establish voluntary pre-release cybersecurity evaluations of frontier AI models with participating labs.
Timeline
- 2026-08-04: White House meets openai, anthropic, google, meta on finalized voluntary framework (2026-08-04-white-house-ai-safety-cyber-testing-meeting)
- 2026-08-03: CNBC/CNN/Reuters confirm ~30-day early access, classified benchmarks, no mandatory licensing (2026-08-04-cnbc-white-house-ai-voluntary-framework)
- 2026-07: Lab self-disclosures — Anthropic cyber eval breaches; OpenAI agent sandbox escape / Hugging Face
- 2026-06: Trump EO directs voluntary cyber testing process
Key Players
Analysis
Formalizes ad-hoc disclosure culture into opt-in government early access. Methodology and public disclosure rules still unpublished. Distinct from eu-ai-act mandatory risk tiers; enterprises face dual-regime reality with ai-governance tooling (asago, astralis).
Related Topics
- ai-cybersecurity-testing
- agentic-ai-regulation
- open-secure-ai-defense
- us-china-ai-competition
- ai-vs-ai-cybersecurity