Obsidian Security has released a new solution designed to enforce real-time controls over AI agents, responding to escalating risks as organizations increasingly rely on these automated systems. The Runtime Policy Decision Engine enables security teams to block, permit, or mandate human oversight for agent actions based on three key factors: the user operating the agent, the agent’s access level, and its interactions with third-party tools. This approach directly addresses a critical vulnerability—many enterprise AI agents operate with broad permissions, creating opportunities for accidental or malicious misuse.
A developer using a coding agent, for instance, could inadvertently expose sensitive credentials or execute destructive commands during a single session. Obsidian’s engine evaluates each action against predefined security policies before execution, ensuring compliance with organizational rules while maintaining operational flexibility.
The system integrates with existing identity frameworks, application permissions, and custom data classifications, such as labels for sensitive files, to deliver precise enforcement. This prevents excessive restriction of legitimate workflows while mitigating high-risk behaviors.
A major obstacle in securing AI agents stems from their distributed nature. These tools often operate across multiple vendors and platforms, each with distinct security models. Obsidian’s solution centralizes oversight by allowing security teams to define policies once, which then apply automatically, whether the agent runs in OpenAI Codex, the Cowork platform, or a custom-built tool. The MCP Security Proxy adds an additional layer of scrutiny for agent-to-application interactions, validating each request before execution. Cloudflare has partnered with Obsidian to extend this control through its MCP Server Portal, ensuring policies remain effective even when agents interact with external services.
Centralized policies stop fragmented AI risks
Avivah Litan, a cybersecurity analyst with Gartner, highlighted the fragmented nature of enterprise agent deployments. Obsidian’s approach resolves this by mapping agent access patterns and enforcing policies across all environments where they operate.
Hasan Imam, Obsidian’s CEO, described the economic stakes of unchecked agent activity. The engine mitigates this risk by applying enforcement rules at the moment of action, before any damage occurs. Centralized policy management reduces manual oversight burdens and contains errors before they escalate.
Read Also: Cybersecurity talent in high demand globally
The engine also extends its scope to AI agents embedded directly within enterprise applications, a category of risk frequently overlooked by security teams. Vendors such as Salesforce, Databricks, and Workday increasingly integrate AI agents into their platforms, sometimes without explicit user consent or security review. These agents often handle sensitive data, including customer records in Snowflake or employee information in Workday, yet their permissions and access patterns remain opaque.
Obsidian’s discovery tools identify these in-app agents, mapping their identities, tools, and data access, then flagging configurations that introduce unnecessary risks. For example, an agent in ServiceNow might automatically trigger workflows affecting production systems without explicit approval for such high-stakes actions. By scanning for embedded agents, the system ensures security policies apply even when originating from third-party code rather than custom-built tools.
Proxy layer secures agent-to-app communications
The solution also closes a critical gap in agent-to-application communication. Many agents interact with business-critical tools, such as Salesforce, Slack, or Google Drive, through MCP servers, which often transmit application tokens directly to the AI client. This creates a security vulnerability: the agent inherits the same permissions as the user without equivalent oversight. Obsidian’s MCP Security Proxy inserts a governed layer between the AI client and supported applications.
Before any tool call executes, the proxy verifies the user’s identity, routes requests to the correct tenant, and ensures the application’s token never leaves the proxy. Each call is then evaluated against Obsidian’s policies, whether to allow, block, or require human approval, using the same rules applied elsewhere in the system. This prevents scenarios where an agent, acting on behalf of a developer, accidentally escalates permissions or accesses restricted files.
The engine’s ability to incorporate customer-defined risk contexts enhances enforcement precision. Unlike generic tools relying on vendor defaults, Obsidian’s policies activate only when aligned with company rules, minimizing false positives while stopping actual threats. For instance, a developer using an agent to review code may legitimately access project files but should never encounter payroll data, even if the agent’s permissions technically permit it. The engine’s context-aware checks enforce these boundaries in real time.
Leave a Reply