Vanta Releases 65 Agent Controls, Drawing a Line Between Governance and Runtime Enforcement
The open-source library gives organizations a shared way to document and assess agent risk, but it does not stop an agent action while it is running.
Listen to this story
The audio brief
Story brief
3 key pointsVanta’s 65-control library gives AI-agent builders and adopters a shared governance baseline, but it stops short of enforcing rules during execution. The split—43 controls for developers and 22 for deploying organizations—covers identity, permissions, memory, vendor oversight, monitoring, and training, with current mappings to ISO 27001 and ISO 42001. Teams will still need identity systems, policy gateways, or...
- 01
Developers get 43 controls; deployers of third-party agents get 22 covering inventories, credentials, vendor reviews, oversight, and training.
- 02
The controls map to ISO 27001 and ISO 42001; SOC 2 and NIST Cybersecurity Framework mappings are planned.
- 03
Vanta’s library can require runtime safeguards but cannot intercept or block an agent action itself.
Vanta has moved Agentic Trust Controls into general availability with 65 open-source controls for organizations that build or deploy AI agents. The release offers a governance checklist for an agent’s access and behavior, while leaving the live enforcement needed to block a prohibited action to other systems.
One library for two roles
The library divides responsibility between the companies creating agents and the organizations adopting agents made by outside vendors. Its design is meant to add agent-specific requirements to existing ISO programs, rather than ask security teams to run a separate framework alongside them.
What the two baselines assign
- Developers receive 43 controls covering agent identity, authority, action guardrails, memory protection, instruction integrity, adversarial testing and runtime instrumentation.
- Organizations deploying third-party agents receive 22 controls covering inventories, credentials, vendor reviews, monitoring, oversight responsibilities and staff training.
Vanta maps the controls to ISO 27001 and ISO 42001, with mappings to SOC 2 and the NIST Cybersecurity Framework planned. That makes the release a translation layer between agent-specific risks and compliance programs organizations already use.
A checklist is not an action gate
The key boundary is operational. Agentic Trust Controls can require restricted tool access, runtime monitoring or human approval, but the library does not itself intercept an agent action and prevent it. Organizations still need technical layers such as identity infrastructure, policy gateways or monitoring software to make those requirements effective.
The market is filling in the second layer
Herman Errico is also working through the Cloud Security Alliance’s Autonomous Action Runtime Management, or AARM, group on a vendor-neutral specification for systems that intercept, authorize and audit agent actions before execution. Its proposed scope includes excessive privileges, prompt injection, irreversible actions and risks that arise across a chain of individually permitted steps.
Commercial products are targeting that enforcement layer now. WitnessAI says its Agentic Control discovers agents and Model Context Protocol servers, applies approved-tool policies and blocks activity at runtime. Those capabilities are complementary to Vanta’s control library: one defines the expected safeguards, while the other category is designed to carry out some of them in production.
The useful test for the library is therefore not whether teams can adopt its checklist, but whether they can produce evidence that an agent’s authority was limited, a policy was evaluated and a prohibited action was blocked. Vanta supplies a common set of questions; the remaining challenge is connecting them to safeguards and audit records.
