How Menlo Security Secures AI Agents from Prompt Injection

Post Views: 7

Organizations are rapidly integrating AI agents such as Microsoft Copilot, Google Gemini, and Claude Code into their workflows without adequate security measures in place, creating vulnerabilities that threat actors are actively exploiting.

Ramin Farassat Discusses Menlo Security’s Approach

Attackers are leveraging prompt injection techniques to embed hidden instructions within web content, documents, and even image files, which AI systems process and execute without human detection. These malicious inputs bypass traditional security controls by targeting the data streams that AI agents interact with rather than the models themselves.

The Threat Landscape

The threat landscape for AI agents involves sophisticated methods where attackers manipulate input channels to inject malicious commands. These techniques can compromise data integrity, execute unauthorized actions, or exfiltrate sensitive information.

Menlo Security’s Solution: MARS

Menlo Security has developed Menlo Agent Runtime Security (MARS) to address this emerging risk. The solution operates by isolating every AI agent session within a secure cloud environment, where content is analyzed and sanitized before being processed by the agent. This approach ensures that potentially harmful inputs are neutralized before they can influence the AI’s behavior.

Key Features of MARS

MARS is now available for deployment, offering enterprises a proactive defense against prompt injection threats. The platform’s architecture prioritizes real-time content analysis, ensuring that AI agents operate within a trusted boundary. This strategy aligns with broader efforts to secure AI ecosystems as their adoption expands across industries.

Expert Insights from Ramin Farassat

Ramin Farassat, Menlo Security’s Chief Product Officer, emphasized that the primary security gaps reside in the connectors and integrations that link AI models to external data sources, rather than the models’ core architectures. Security teams must implement controls that monitor and restrict the attack surface of agentic AI systems without hindering their functionality.

“By focusing on the environments where AI agents interact with data, organizations can mitigate risks while maintaining operational efficiency.”

Availability and Events

Menlo Security is showcasing the solution at Black Hat USA (Booth #4702, Mandalay Bay) and Ai4 (Booth #424, The Venetian) from August 4–6 in Las Vegas. The platform’s architecture prioritizes real-time content analysis, ensuring that AI agents operate within a trusted boundary.