The control gap is now operational.
Models improved. Agents gained more room to act. Governance expectations hardened. That is why evidence can no longer begin after the action.
For serious buyers and partners, the live question is no longer only what AI can do. It is whether review, provenance, testing, and accountability arrive before consequential use.
Why now is no longer abstract.
The pressure is not coming from one place. It is coming from capability progress, real deployment, uneven safeguards, and rules that now have dates.
Review windows are compressing.
AISI says performance in some tested areas is doubling roughly every eight months, and frontier models can now complete software tasks that would take a human expert over an hour.
AI is moving closer to consequential action.
AISI reports increasing deployment in critical sectors and a rise in higher-autonomy finance-focused agent tooling, including systems positioned closer to asset transfer and trading operations.
Improvement is not the same as hard control.
AISI found universal jailbreaks for every system it tested. Safeguards can improve, but post-hoc monitoring is still not a substitute for dependable pre-action controls.
Governance now arrives with dates and workflow duties.
The EU AI Act now applies progressively, OMB has issued use and acquisition memos, and enforcement and high-risk obligations continue to move into view.
Recent dates matter here.
For serious buyers and partners, the external timeline makes the shift concrete.
EU AI Act first provisions apply.
General provisions, AI literacy, and prohibitions start to apply.
OMB issues new AI use and acquisition memos.
Federal AI use and buying now sit inside named governance and cross-functional workflows.
EU GPAI rules apply.
Rules for general-purpose AI apply and governance must be in place.
AISI publishes its first public frontier trends report.
The macro picture becomes measurable, not just anecdotal.
Major EU AI Act rules and enforcement start.
High-risk Annex III rules, Article 50 transparency rules, and enforcement begin.
The baseline is getting clearer.
Serious review increasingly expects four things early: a governance owner, pre-deployment testing, provenance and documentation, and a cross-functional review path.
Owner
Someone identifiable has to own the governance decision, not just the deployment enthusiasm.
Test
Pre-deployment testing and risk mitigation need to reflect expected real-world outcomes, not only demo conditions.
Provenance
Content, model, and system facts need enough traceability and documentation to survive internal review.
Review
Security, legal, procurement, and technical teams need a shared way to assess risk before consequential use scales.
Primary sources behind this page.
This page is built on primary sources rather than commentary. Official materials carry the case.
Frontier AI Trends Report
Rapid capability growth, uneven safeguards, social effects, and signs of more autonomous activity in critical sectors.
Open sourceM‑25‑21
AI use guidance covering governance, high-impact AI, minimum risk management practices, and pre-deployment testing.
Open sourceM‑25‑22
AI acquisition guidance emphasizing cross-functional teams, vendor transparency, and procurement discipline.
Open sourceAI 600‑1 GenAI Profile
Governance, content provenance, pre-deployment testing, and incident disclosure as practical risk-management themes.
Open sourceEU AI Act timeline
The progressive application calendar that makes the regulatory side of “why now” concrete.
Open source