On 9 November 2026, Vanta announced a definitive agreement to acquire Pillar Security, a young company that tests AI agents for weaknesses and protects them while they run. Vanta is best known for compliance automation: it collects evidence for audits such as SOC 2 and ISO 27001 and tracks controls across hundreds of connected systems. With Pillar, it wants to go a step further and act on AI agents directly, not only report on them.
What changed
Pillar was founded in 2023 by Ziv Karliner and Dor Sarig. Its products cover three areas: adversarial testing of AI applications and agents (often called red teaming), monitoring of agent behaviour at runtime, and blocking actions that break policy. Its research team also wrote SAIL, a framework for securing AI agents that maps to ISO and NIST standards and to OWASP guidance.
Vanta says it will connect these capabilities to its Trust Graph, the model of a company's security and compliance programme that it builds from more than 400 integrations. The idea is that a governance rule, such as "this agent may not export customer records", stops being a line in a policy document and becomes a boundary the platform can enforce. If an agent tries something outside its permissions, the action is blocked, the Vanta Agent alerts the owner and an audit record is created.
Vanta has published a staged plan for 2027:
- January 2027: an AI inventory that lists AI applications, agents, models and MCP servers, linked to owners, data access, vendors, risks and controls. Runtime monitoring and detection arrive at the same time.
- Spring 2027: red teaming of AI systems.
- Summer 2027: runtime enforcement.
Vanta also says AI Governance will become a paid offering from January 2027. No price for the deal was given by the companies. Press reports put it at around 150 million US dollars, but neither side has confirmed that figure, and the closing date has not been stated.
Why it matters
Most organisations now have AI agents in more places than their security teams can name. Some are built in-house on cloud AI platforms, some arrive inside SaaS products, and some are set up by staff with a browser and a free account. Each one can read data and take actions, often with credentials that were never reviewed.
Compliance tools have been good at answering "is the control in place?" once a quarter. They have not been built to watch an agent decide, in the moment, to call a tool or move a file. Bringing runtime protection into the same place as the control framework closes part of that gap. The evidence that an agent stayed within its limits can feed straight into an audit, rather than being rebuilt by hand from logs.
There are practical points to weigh. Runtime enforcement is the last item on the plan and is not due until summer 2027, so for most of next year this will be visibility and detection, not blocking. Teams should also plan how a new runtime layer will sit alongside controls they already run, such as AI gateways, data loss prevention and identity tools for non-human accounts. Overlapping policies in two places is a common source of gaps.
For existing Pillar users, Vanta has not yet said how the standalone product will be handled. That is worth asking about before renewal.
What this means for organisations in the Middle East, India, Africa and North America
Regulators in our markets are starting to ask for proof of how AI is used, alongside a written policy. In the UAE and Saudi Arabia, personal data protection laws expect firms to know which systems process personal data and on what basis. India's DPDP Act puts similar duties on data fiduciaries. In Canada, OSFI's guideline E-23 on model risk takes effect for federally regulated financial institutions in 2027 and covers AI models. In the USA, many firms map their AI controls to the NIST AI Risk Management Framework.
All of these start with the same question: what AI do you have, who owns it and what data can it touch? The AI inventory Vanta plans for January answers that question directly, and it is the part of the roadmap most teams can use first. Linking each agent to an owner and a control also makes it easier to show a regulator or a customer auditor that someone is accountable.
Data residency is another factor. Before sending agent prompts, outputs or logs to a new monitoring service, check where that data will be stored and processed, and whether that fits sector rules for banking, health or government work in your country.
What to do next
- Build an AI inventory now. List agents, models and MCP servers in use, with an owner, the data each one can reach and the accounts it uses. You will need this whatever tool you pick.
- Write agent boundaries as testable rules. For each agent, state what it may and may not do in plain terms that a tool could later enforce.
- Plan for monitoring before blocking. Since enforcement arrives last, decide how alerts will be triaged during 2027 and who can switch an agent off.
- Ask about cost and overlap. If you already use Vanta, ask how AI Governance will be priced and how it will work with your existing AI gateway and identity controls.
- Run a red-team exercise on one agent. Test a high-risk agent against prompt injection and data leakage before the spring 2027 tooling arrives, so you have a baseline.
You can read more about how we work with Vanta on our Vanta partner page, and about our approach to AI governance.