Menlo Security unveils MARS to defend against emerging prompt injection threats in AI agents

As enterprises rapidly deploy AI agents like Microsoft Copilot and Google Gemini, Menlo Security introduces MARS, a cloud-based runtime to safeguard against hidden prompt injections and web-based attacks, addressing the security gaps in autonomous AI adoption.

Enterprises are rolling out AI agents such as Microsoft Copilot, Google Gemini and Claude Code faster than they are building the controls to secure them, creating an opening that attackers are exploiting through prompt injection. Menlo Security says the danger is not always in the model itself, but in the content and connections around it, where hidden instructions can be buried in web pages, files and even images in ways a human would miss but an agent may still follow.

To address that gap, the company has developed Menlo Agent Runtime Security, or MARS, a cloud-based runtime designed to place every agent session inside an isolated environment. According to Menlo Security, the system strips malicious instructions from content before the agent can act on it, with the aim of reducing the risk of invisible prompt injections, logic attacks and data leakage.

Ramin Farassat, Menlo Security’s chief product officer, argues that security teams need to focus on the agent attack surface rather than trying to block agentic AI altogether. The company says MARS also helps bridge the “API gap” by letting agents reach legacy systems that lack modern interfaces, while applying controls over what data they can see and where it can go.

Menlo Security’s materials also describe the product as running browser sessions in remote, disposable containers, sanitising files and pages, and masking sensitive information at the source to reduce exfiltration risk. The broader pitch is that enterprises should be able to expand the use of autonomous tools without exposing themselves to web-based threats that are difficult for people to spot but simple for machines to process.