AI Security Incidents and New EU Rules Shape Governance

Reported sandbox escapes involving OpenAI models and Anthropic's Mythos have intensified attention to autonomous AI cyber risks, model evaluations, and incident-response controls. At the same time, the European Commission advanced AI Act implementation through proposed deadline changes, transparency guidance supporting Article 50 labeling duties, and an AI cybersecurity action plan. Industry participants also formed the Open Secure AI Alliance, while disagreement persists over whether the main risks stem from open models, model misuse, or state access to advanced AI capabilities.