Vanta has signed a definitive agreement to buy Pillar Security, a two-year-old startup that tests and guards AI agents while they run. Price and terms weren’t disclosed. The same day, Orca Security put out its 2026 State of AI Security Report, and one number in it explains why a compliance company wants a runtime security team: of the AI vulnerabilities that already have a fix, 99.9% are still sitting unpatched.
Read together, the two releases say the same thing from opposite ends. Companies have written their AI policies. They have not enforced them.
What Vanta is buying
Pillar was founded in 2023 by Ziv Karliner, its CEO, and Dor Sarig. It sells a platform that finds AI agents inside a company, red-teams them the way an attacker would, and watches and blocks what they do in production. Vanta built its business on the other half of the problem: turning SOC 2, ISO 27001 and similar frameworks into automated evidence collection. It says it passed $300 million in annual recurring revenue in April 2026, growing 63% a year, and now has more than 20,000 customers, up from 16,000 in the first quarter. Three of every four companies on the Forbes AI 50 list use it.
Vanta laid out a schedule for folding Pillar in. An AI inventory and runtime monitoring and detection arrive in January 2027. Red-teaming follows in spring. Runtime enforcement, the piece that actually stops an agent from doing something, comes in summer 2027. This is Vanta’s third acquisition, after Trustpage in January 2023 and Riskey in July 2025, and the first that moves it from proving controls exist to operating them.
Christina Cacioppo, Vanta’s CEO, put the logic bluntly. A policy document alone doesn’t get a company to safe AI. IDC analyst Phil Harris framed it as the market moving from governance on paper to governance at runtime.
The Orca numbers behind the deal
Orca drew its report from anonymized telemetry across more than 1,200 production cloud environments on AWS, Azure and Google Cloud, collected in the second quarter of 2026. A few findings stand out.
- 56% of organizations already run AI agents in production, and 51% use AI to build their own applications.
- 81% of organizations using AI packages have at least one known vulnerability in them, up from 62% in Orca’s 2024 report.
- Half of those AI package vulnerabilities now have public exploits.
- 64% run vector databases, and 55% run four or more AI services at once.
- Between 87% and 98% of AI workloads on the three big clouds don’t use customer-managed encryption keys.
There is some good news. The share of Amazon SageMaker environments running with root access fell from 98% to 76%, and insecure instance metadata settings dropped from 77% to 48%. Those are configuration fixes a team can make once. Patching AI libraries is different. It breaks things, the dependency trees are deep, and the packages ship new versions every few weeks. That is why the fix rate is near zero.
Why compliance vendors are moving into runtime
The regulatory calendar is doing some of the pushing. The EU AI Act’s high-risk obligations took effect on August 2, 2026, and Colorado’s amended AI Act starts on January 1, 2027. Auditors under both will want proof that controls work in production, not just a written policy and a training log.
That puts compliance platforms like Vanta, Drata and Secureframe on a collision course with security vendors that already sit in the runtime path. Vanta’s bet is that customers would rather extend the tool their auditors already accept than add another console. Whether Pillar’s enforcement layer can run at the scale of 20,000 customers, most of them small companies with no AI security staff, is the open question for 2027.
For buyers, the Orca data is the more useful read. If an agent can reach company data and the libraries under it carry known, exploitable bugs, the policy that says it shouldn’t misbehave is not a control.