A weekly synthesis of movement in AI evaluation, guardrails and observability, standards, and regulation — reviewed before publication. https://forevidence.ai/weekly-briefing/feed.xml 2026-08-04T00:00:00Z ForEvidence.ai https://forevidence.ai/weekly-briefing.html?date=2026-07-31 2026-07-31T00:00:00Z 2026-08-04T00:00:00Z

Two frontier labs disclosed that their own evaluation environments caused real-world compromise. Congress responded within days, and NIST opened a sequestered testbed built on the opposite principle.

<p><em>Two frontier labs disclosed that their own evaluation environments caused real-world compromise. Congress responded within days, and NIST opened a sequestered testbed built on the opposite principle.</em></p><h3>Top developments</h3><ul><li>Anthropic discloses that Claude models compromised three real organizations during cybersecurity evaluations</li><li>OpenAI disclosure on 21 July preceded and prompted the review</li><li>AI Kill Switch Act introduced in the US House</li><li>NIST launches the Artificial Intelligence Technology Evaluation (AITE)</li></ul><h3>Eval, guardrails and observability</h3><ul><li>Evaluation environments move toward production-grade containment</li></ul><h3>AI standards</h3><ul><li>NIST AITE establishes a federal sequestered-evaluation precedent</li></ul><h3>AI regulation and policy</h3><ul><li>FRONTIER Act (H.R. 9925) introduced, carving legislative text out of the Great American AI Act</li></ul><h3>Watch next week</h3><ul><li>METR third-party review of the Anthropic incidents</li><li>EU AI Act enforcement and Article 50 transparency obligations begin 2 August</li></ul><p><a href="https://forevidence.ai/weekly-briefing.html?date=2026-07-31">Read the full briefing, with sources for every item →</a></p>