Obsidian Security has announced a runtime policy engine designed to allow, block or require approval for AI agents’ actions across business applications. The useful shift is from spotting agents and their access to applying controls at the moment they act, rather than trusting that yesterday’s permissions will behave today.
Watch Desk analysis
What happened
Obsidian’s engine draws on identity, employment, application permissions and customer-defined context, such as data-classification labels, to assess actions against centrally managed policies. Connected enforcement points can allow or block an action, or require human approval where supported, according to Help Net Security’s report of the announcement.
The company also announced runtime policy support for OpenAI Codex, alongside Claude Code and Cowork, and discovery and governance for agents built into platforms including Salesforce Agentforce, Databricks, GitHub, Snowflake, ServiceNow and Workday. Its new MCP Security Proxy checks agent tool calls before they reach supported applications. A Policy Decision API is intended to let customer-managed gateways and custom agent tools request decisions from Obsidian’s engine.
Why it matters
An agent can have legitimate credentials and still do something its owner did not intend. Obsidian’s pitch is to judge the action in context: who is behind the agent, what it can reach and, where configured, how sensitive the data is. That could give organisations a common policy layer across coding tools, workplace apps and custom gateways, instead of a growing collection of rules that do not speak to one another.
The practical promise is meaningful, but it is a vendor announcement, not evidence yet of how well these controls work in real deployments. “Require approval” also depends on which enforcement points support it and how organisations configure them. A policy engine is only as useful as the context and coverage it actually has.
Our read
This is a substantial, specific answer to a real AI security problem: agents need access to do useful work, but access should not be a blank cheque. The breadth of the announced controls is more interesting than another dashboard that merely counts agents. Security teams will want to test how consistently policies follow actions across tools, and how much friction they add to ordinary work.
What to watch
- Which applications and enforcement points support blocking or human approval at launch.
- How the MCP Security Proxy and Policy Decision API become available to customers.
- Whether organisations can apply consistent policies across third-party and in-house agents.
- How often policy decisions interrupt legitimate work, as well as risky actions.
Discussion spark: Should AI agents be allowed to act freely within approved access, or should consequential actions require a person’s approval?
Sources and evidence
- Source update (11 November 2026, 08:44 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.