August 10, 2026 — curated links and takeaways.
1. OpenAI and Anthropic Cybersecurity Incidents Expose Safety Testing Risks
OpenAI disabled safety restrictions on Astra during cyber evaluations; Anthropic ran tests without standard public safeguards, leading both companies to pause work pending new containment protocols. Builders evaluating frontier models need explicit clarity on which safety constraints are disabled during adversarial testing and formal audit trails of such periods.
2. The AI Safety Test Is Becoming a Safety Risk
Anthropic's postmortem on three incidents revealed inadequate monitoring during safety evaluations—clear warning signs were missed by both Anthropic and testing partner Irregular. Teams deploying custom safety evaluations must implement independent oversight and real-time anomaly detection to avoid repeating this pattern.
3. Munich Court Rules Suno Liable for Copyright Infringement in Music AI Training
Munich Regional Court I ruled July 31 that Suno infringed GEMA-represented copyrights during training and in outputs, despite training occurring in the US—establishing extraterritorial enforcement precedent. AI builders in music, audio, and text generation now face material risk of injunctions and damages under EU/German law regardless of where training occurs.
4. EU AI Act Compliance Deadlines Hit August 2026: Marking, Transparency, and Fines Active
As of August 2026, EU AI Act prohibitions, transparency mandates for generative AI, and high-risk system requirements are enforceable with fines up to €35M or 7% global turnover for prohibited practices. Compliance teams must audit training data sources, implement output disclosure mechanisms, and document risk assessments before December 2, 2026 marking deadline.