AI Safety and the New Governance Landscape

As AI capability accelerates, the systems meant to keep it in check are under pressure. The Stanford AI Index found that responsible AI is “not keeping pace with capability,” with safety benchmarks lagging and documented incidents rising to 362 in a year.

The EU leads with binding law

The EU AI Act and the newer AI Omnibus represent the world’s most ambitious effort to regulate AI. They forbid certain applications and impose transparency and oversight requirements on high-risk systems and foundation models.

Critical capability thresholds emerge

OpenAI designated GPT-6 Astra as its first model to cross a “critical cybersecurity capability threshold” — meaning it can find and exploit previously unknown vulnerabilities across well-protected systems. That designation triggers stricter access limits and monitoring.

The alignment problem

OpenAI’s chief scientist noted that “progress in intelligence does not guarantee progress in alignment,” and that monitoring AI reasoning is getting harder as models use fewer language tokens.

The industry response includes:

• Misalignment monitoring — 24/7 systems that inspect model reasoning for signs it’s operating outside its authority
• Red-team testing — including a growing role for government review before release
• Restricted access — gating the most powerful capabilities behind trust and security vetting

The governance question is no longer whether to regulate AI, but how to regulate it fast enough to keep pace with the technology itself.