Obsidian launches runtime policy engine to secure AI agents
The new tool allows security teams to centrally manage and block specific actions taken by AI agents across various enterprise platforms.
The short version
- The engine evaluates agent actions against central policies before they reach the target application.
- It supports major platforms including Microsoft Copilot, Salesforce Agentforce, and Snowflake.
- Security teams can use custom data labels to define what is sensitive for specific policy rules.
- A new MCP Security Proxy monitors tool calls to keep app tokens out of the AI client's hands.
Why it matters
As companies adopt AI agents that can perform tasks like writing code or moving data, they face risks from agents acting with too much authority. This tool helps organizations give agents the access they need to be useful while preventing them from performing high-risk or unauthorized actions.
The challenge of agent authority
Enterprises are increasingly using AI agents to automate workflows, write code, and manage data. While these tools increase efficiency, they also introduce significant security risks because agents often operate with the credentials of the humans who deploy them. An agent might have the legitimate permission to view customer records but could still be used to perform an action that causes massive damage.
Security teams currently struggle to keep up with the speed at which these agents are being integrated into products. Many third-party vendors are embedding agents directly into popular software like Workday, ServiceNow, and Snowflake. These agents often arrive with features already switched on, frequently bypassing traditional security reviews and sitting close to sensitive business data.
Obsidian Security is addressing this by launching a Runtime Policy Decision Engine. This system is designed to give security teams centralized control over what an agent can do at the exact moment it tries to perform an action. It aims to move defense away from static rules toward a model that understands the context of every individual request.
How the policy engine works
The new engine works by gathering context from multiple sources before an agent is allowed to complete a task. It looks at who is behind the agent, what third-party applications are involved, and what specific permissions are being requested. This information is combined with the company's own internal data, such as identity systems and employment records.
One of the most important features is the ability for companies to provide their own context, such as data classification labels. For example, a security team can define a rule that blocks agents from accessing files labeled as highly sensitive unless the user is part of the executive team. This allows the policy to be precise rather than just blocking all AI activity across the board.
When an agent attempts a tool call, the system evaluates the request against these centrally managed policies. Depending on the outcome, the engine can allow the action, block it entirely, or mandate that a human must approve the request first. This ensures that high-risk actions are caught while routine, low-risk tasks can proceed without friction.
Expanding coverage to coding tools
Developers are adopting coding agents at a rapid pace, often using multiple different vendors simultaneously. These agents can run with access to source code, databases, and production systems, which creates risky moments during a session. For instance, an agent might accidentally paste a secret into a prompt or execute a destructive command on a server.
Obsidian is expanding its runtime policy enforcement to include OpenAI Codex, joining existing support for Claude Code, Cowork, and Microsoft Copilot. This means security teams can manage the safety of various coding agents from a single console. They can ensure that developers keep their preferred tools while the company maintains oversight of the underlying actions.
By providing a unified console, the platform helps prevent a fragmented security posture. Instead of writing different rules for every different coding tool, administrators can define a policy once. This approach helps contain mistakes and allows the business to move faster on AI adoption without sacrificing safety.
Proxying and gateway integration
The company also introduced an MCP Security Proxy to handle specific types of agent interactions. Many agents reach applications like Slack or Google Drive through MCP servers, sometimes even handing the AI client a direct app token. The new proxy sits between the AI client and the app to identify the user and keep the app's token hidden from the AI client.
Additionally, Obsidian is offering a Policy Decision API for customers who already use their own gateways. Many enterprises route agent traffic through custom API or AI gateways, but these gateways often lack the context of who is behind a request. The new API allows these gateways to send each action to Obsidian's engine to receive a decision based on identity and data labels.
Obsidian is also partnering with Cloudflare to provide contextualized policies for agentic access. By connecting Cloudflare to the Policy Decision API, customers can apply informed decisions at the point where traffic already flows. This allows for stronger control over agent activity without requiring the company to rebuild its existing infrastructure.
The future of agent governance
The goal of these new tools is to change the economics of security for AI. As Hasan Imam, CEO of Obsidian Security, noted, an agent can hold valid credentials and still take an action that costs a business millions. By mapping the full chain of authority across the third-party estate, the company aims to make defense cheaper and more reliable.
The expansion of discovery and governance capabilities also covers major platforms like Salesforce Agentforce and Databricks. This helps security teams identify in-app agents that might be built by users without a formal review. Mapping these identities and tools helps close a gap where agents act across different SaaS applications and data platforms.
Moving forward, the focus is on creating a single policy layer that follows agents into every application they touch. By using a centralized engine that understands the value of the data and the identity of the user, companies can manage the risks of autonomous agents. This setup ensures that as AI becomes more integrated into the workplace, the security measures scale alongside it.
Key takeaways
Why it matters
As AI agents gain the ability to perform complex tasks autonomously, they become a significant security liability if they have unrestricted access to company data. This new engine provides a way for businesses to set
What to watch
- How many other major SaaS platforms will integrate with the Obsidian Policy Decision API?
- Whether more companies will adopt the MCP Security Proxy to hide app tokens from AI clients.
Try it
Check if your organization uses any AI agents or 'copilots' that have access to your company's internal databases or file systems.
Sources and links
First reported by Help Net Security. 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.