Frontier-model governance is becoming an evaluation problem before it becomes a licensing regime. A White House order creates federal cyber-testing infrastructure without mandatory preclearance, Anthropic is widening controlled Mythos access, and LongJudgeBench finds that model graders remain unstable on document-scale work.
1. White House order creates a voluntary frontier-model review path
Executive Order 14409 directs federal bodies to prioritize cyber defense, form an AI cybersecurity clearinghouse, and develop a classified benchmark for identifying covered frontier models. It also calls for a voluntary process through which developers can provide federal access before broader trusted-partner release.
The pre-release period can last up to 30 days, while the order explicitly rejects mandatory licensing, preclearance, and permits. Its immediate force lies in building a classified benchmark and secure federal evaluation capacity before participation has legal compulsion.
Sources: White House Executive Order 14409
2. Project Glasswing expands controlled access to Mythos Preview
Anthropic said it is adding approximately 150 organizations across more than 15 countries to Project Glasswing. The new cohort includes power, water, healthcare, communications, hardware, and widely used software vendors, following an initial group of roughly 50 partners.
Anthropic says existing safeguards cannot support unrestricted Mythos access, so participation carries security conditions. Expanding from roughly 50 to about 200 organizations tests whether coordinated disclosure and patching can absorb more model-generated findings without opening the same capability broadly.
Sources: Anthropic's Project Glasswing expansion
3. LongJudgeBench finds long-form model grading remains unstable
LongJudgeBench evaluates LLM judges across long-form scenarios requiring document-level assessments of organization, coverage, consistency, depth, and task-specific quality. The paper reports substantial variation across scenarios and judging setups, with rubrics and reference answers helping but not reliably closing the gap.
The instability makes a single model-generated score especially fragile for reports, research synthesis, and multi-file code work. Rubrics narrow some variance, yet the remaining scenario sensitivity can convert a judge model's preferences into apparent measurement.
Sources: LongJudgeBench paper · LongJudgeBench repository