On April 8, 2026, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell called the CEOs of Citigroup, Morgan Stanley, Bank of America, Wells Fargo and Goldman Sachs to Treasury headquarters to discuss the cybersecurity risk posed by a model that had not even shipped yet: Anthropic's Claude Mythos Preview, according to CNBC. That meeting is the clearest marker of what actually changed in enterprise AI vendor security review this year. Review did not get more thorough across the board. It split in two.
A two-tier system, not a better one
At the top, a small number of frontier AI labs now receive scrutiny once reserved for systemically important financial infrastructure. Within days of the Treasury meeting, Anthropic answered with Project Glasswing, giving Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks early access to Claude Mythos Preview to find and patch vulnerabilities before general release, backed by up to $100 million in model-usage credits. That is a vendor standing up a private, named-counterparty, pre-release audit program because a systemic-risk conversation forced its hand.
Everywhere else in the enterprise, review has not kept pace with exposure. Vanta's State of Trust 2025 report, drawn from more than 16,000 customers, found roughly 70% have shadow AI usage inside their organization, and that while 30% of the AI vendors in use are rated critical or high risk, only 7% of those vendors are actually reviewed. Buyers are identifying risk faster than they are acting on it.
The infrastructure hasn't caught up
The reviewing muscle itself is still underbuilt. KPMG's 2026 Global Third-Party Risk Management Survey of 851 organizations found only 53% of third-party risk programs are "mostly integrated" with enterprise risk management, and just 18% are "fully integrated," even as AI vendor exposure grows. Inside the security function specifically, it is no better: KPMG's 2026 Cybersecurity and Technology Risk Survey of 310 security leaders at organizations with $1 billion or more in revenue found only 24% say AI is fully integrated into their cybersecurity program, while 53% report only partial integration.
Policy lags furthest behind sentiment. Darktrace's State of AI Cybersecurity 2026 report, surveying 1,540 security leaders across 14 countries, found just 37% of organizations have a formal AI policy in place even as 92% say they are concerned about AI agent security risk. Buyers know what worries them. Most have not written down what they will do about it.
Concern is universal. Review is not. That is the gap enterprise buyers are actually operating inside right now.
Why review is about to get forced anyway
The gap will not stay optional much longer. The EU AI Act's remaining provisions become legally binding on August 2, 2026, including high-risk system obligations for providers under Articles 9 through 17 and deployer obligations under Article 26, meaning any enterprise buying AI tools for EU operations now needs documentation and audit-readiness from vendors, not just an internal comfort level. That deadline is doing what internal risk committees have not: setting a hard date for review to become mandatory rather than aspirational.
The vendors themselves are giving buyers reason to move faster. Black Kite's 2026 Supply Chain Vulnerability Report counted 2,130 AI-related CVEs published in 2025, a 34.6% year-over-year increase, with GitHub Copilot, Cursor and Claude Code recording their first high-severity CVEs. These are not obscure vendors. They are tools already embedded in enterprise developer workflows, shipping vulnerabilities at an accelerating rate while, per Vanta, only 7% of the AI vendors enterprises rate critical or high risk are actually getting reviewed.
What this means for a buyer this week
The change in 2026 is not that security review requirements got harder across the board. It's that they got harder for a small number of vendors deemed systemically important enough to summon bank CEOs to Treasury, while the long tail of AI tools quietly accumulating inside every other enterprise still isn't getting checked. An operator does not need to wait for a Treasury-level crisis to close that gap. Pull the vendor list this week, flag every AI tool that would rate critical or high risk under Vanta's method, and push those specific vendors through review before August 2 turns the absence of documentation into a compliance finding instead of an internal risk decision.