SIAgentNews.com
← All news

Obsidian Security launches runtime policy engine for AI agents

The new tool allows companies to monitor and block specific actions taken by AI agents in business apps, even when the agent has valid credentials.

· First reported by Unite.AI

The short version

  • The engine checks if an agent's action is appropriate based on the person behind it and the specific data being accessed.
  • Supported platforms include Salesforce Agentforce, Databricks, GitHub, Microsoft 365, Snowflake, ServiceNow, and Workday.
  • A new MCP Security Proxy prevents AI clients from storing direct application tokens, using proxy-issued tokens instead.
  • The system includes a 'flag mode' to test how policies would affect real traffic before actually blocking any actions.
  • It provides specific controls for coding tools, including support for OpenAI Codex.

Why it matters

As companies give AI agents more power to perform tasks, standard permissions are often too broad to prevent mistakes. This technology provides a way to let agents do their jobs without giving them the power to perform unauthorized or destructive actions.

The limits of standard access

Enterprises face a unique security hurdle when deploying AI agents. Traditional access controls often fail because an agent might hold valid credentials to an application but still attempt an inappropriate action. For example, an agent might have permission to read customer records but should not have the power to export them to an external destination.

Obsidian Security is addressing this gap with its new Runtime Policy Decision Engine. This tool is designed to evaluate what an otherwise authorized agent is allowed to do at any specific moment. It moves beyond simple 'yes' or 'no' permissions to look at the context of the request.

As reported by Unite.AI, the engine analyzes the person behind the agent, their employment status, and their specific permissions. It also considers the application involved and any customer-supplied data classifications to decide if a request should proceed.

Context-aware security architecture

The system relies on a Knowledge Graph architecture to provide the necessary context for these decisions. This architecture connects various data points including people, accounts, applications, tokens, and entitlements. By normalizing information from administrative APIs and activity streams, the system can trace access across multiple systems instead of looking at accounts in isolation.

This structure helps the policy engine understand complex relationships. For instance, an agent might appear to have a narrow role, but it could inherit broader reach through a specific token or a shared entitlement. The Knowledge Graph allows the engine to see these connections to make more accurate security decisions.

Customers can also inject their own specific context into the system. This includes adding their own classification labels to files. This allows a company to distinguish between an ordinary document and one the business designates as highly sensitive, rather than relying on a generic definition of confidential data.

How enforcement works

Obsidian separates the process of deciding on an action from the act of enforcing it. When an enforcement point pauses a proposed operation, it submits the details to the engine for evaluation. The engine then returns a decision and a reason, which the enforcement component applies before the operation is allowed to proceed.

The company released a Policy Decision API to extend this model to various customer-operated gateways and workflows. This means an enterprise can keep its existing infrastructure while using Obsidian to provide context-aware verdicts. The system also includes a flag mode that lets teams test policy matches against real traffic without actually blocking anything.

For those who need human oversight, the system supports approval workflows in specific cases. Currently, the company's documentation notes that requiring manual approval is available for Claude Code policies that run before an operation. This allows for a mix of automated blocks and human reviews depending on the tool being used.

Protecting tokens with MCP Proxy

The Model Context Protocol (MCP) provides a standard way for agents to call tools from external servers. To secure this, Obsidian launched the MCP Security Proxy. This proxy sits between the AI client and the application's MCP server to govern the connection and prevent the AI client from storing direct application access tokens.

In this setup, users authorize access through the standard vendor consent flow. The AI client then receives a proxy-issued token instead of the actual application token. The proxy maps each session to the specific user, their application account, and the relevant tenant, such as a specific Salesforce organization or workspace.

This proxy-based approach ensures that the agent operates within the user’s existing permissions. It does not add new privileges through a shared service account. Additionally, Obsidian is collaborating with Cloudflare and its MCP Server Portal to combine gateway positioning with identity and data context.

Broad coverage for enterprises

The new engine is designed to cover a wide range of enterprise software and coding tools. It includes support for major platforms like Databricks, GitHub, Microsoft 365, Snowflake, ServiceNow, and Workday. It also specifically covers Salesforce Agentforce, which is a key platform for deploying business agents.

Beyond standard business apps, the tool provides coverage for coding agents. This includes support for OpenAI Codex alongside existing controls for other coding-agent tools. These controls are intended to prevent risks like destructive commands, credential exposure, or reads that fall outside of a project’s intended scope.

The goal of these action-level controls is to balance utility with safety. A total ban on agents can stop a workflow entirely, while a narrowly scoped decision allows routine tasks to continue. By bringing these decisions into a shared policy layer, companies can let agents perform useful work while containing potential mistakes.

Next steps for security teams

Security teams can use the system to gain better visibility into agent behavior. The architecture captures decision records that include the agent, the invoking person, the tool used, the arguments provided, and the final outcome. This provides a clear audit trail for investigating why an action was allowed or stopped.

As agents gain more operational responsibility, the ability to explain and enforce a decision before it happens becomes vital. The system is designed to help organizations move from simple discovery—knowing what tools are present—to active governance and runtime enforcement. This ensures that agents remain within the boundaries of company policy.

The value of this rollout will depend on how well a company can map its internal permissions and data classifications. Organizations will need to ensure their Knowledge Graph is populated with accurate information to get the most out of the policy engine's context-aware decisions.

Sources and links

First reported by Unite.AI. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news