This page may contain stale information. Last updated: 2026-08-14
Definition
- 2026-08-03: horizon3 2B+ for autonomous-penetration-testing (2026-08-03-horizon3-series-e-official)
Voluntary U.S. federal framework for pre-release review of frontier AI models for advanced cyber capabilities.
Key Points
-
2026-08-14: Vendor CyberGym/ExploitBench claims for GLM-5.3; treat as unverified (cybergym)
-
2026-08-05: aisi lessons — fine-grained network controls, real-time monitoring, justify internet access; not sandbox escape (2026-08-05-aisi-unsanctioned-agent-cyber-testing)
-
2026-08-04: White House meets openai, anthropic, google, meta to present finalized voluntary cyber safety tests; ~30-day early access; classified benchmarks; not mandatory licensing (2026-08-04-cnbc-white-house-ai-voluntary-framework, us-voluntary-ai-cyber-testing)
-
2026-07-30: Lab self-disclosure — Claude cyber evals caused three production breaches + pypi-malware; prompts OpenAI-style review culture (2026-07-31-anthropic-claude-cyber-evals-three-breaches, frontier-lab-eval-safety)
-
2026-07-14: gold-eagle clearinghouse operationalizes EO 14409 vulnerability coordination (2026-07-18-white-house-gold-eagle-official)
-
2026-07-15: GPT-Red case study compromised live Codex CLI and office vending agent (Vendy) before disclosure (2026-07-16-openai-gpt-red-official-primary)
-
Trump EO signed June 2, 2026
-
Up to 30-day voluntary review before public release
-
Treasury-led AI cybersecurity clearinghouse
-
Classified benchmarking defines covered frontier models
-
Explicitly not mandatory licensing