
An AI agent handling a refund request starts a funds transfer that no one asked for and no one approved. Would the company notice before the money moved?
This is the problem Lasso is addressing. Part of the NetApp Excellerator program, it protects the AI agents enterprises build and use, wherever they run, by learning what each agent is meant to do and checking every action it takes against that purpose.
Enterprises are putting AI agents to work faster than they can keep track of them. These agents plan their own steps, call tools and MCP servers on their own, and act without a person approving each move. That autonomy is the reason companies deploy them, and it is also how an agent ends up outside the boundaries it was given.
Two challenges make this hard to control.
The challenges of securing AI agents
Agents are everywhere, and each one has its own purpose. AI agents now run in every part of the enterprise. Developers use coding assistants on their endpoints and connect them to MCP servers and skills. Engineering teams build custom agents that run in code and in the cloud, and business teams create their own agent using low-code no-code builder platforms.
Each of these agents are set up for a different purpose, and each one works toward it autonomously. With this many agents running in this many places, it becomes difficult to track what any single agent is meant to do, and without that there is no reliable way to tell when one has gone off task or gone rogue.
Context windows are large, and attacks hide inside them
Each agent gathers the context it needs for a task by calling models, MCP servers, tools, and data sources, and everything it gathers goes into its context window. For a single task, that context can reach hundreds of thousands of tokens. A prompt injection can be one sentence hidden in an email, a web page, or a tool response, and inside a context that large it becomes the needle in the haystack. Most detection models were built for short prompts. On inputs of this size, they slow down or miss the attack, and an agent that reads the injected instruction may act on it as if it came from the user.
How Lasso does it
Lasso’s AI security platform protects the AI agents enterprises build and use, wherever they run, whether that is on an endpoint, in code or in the cloud.
It works from each agent’s intent. Its Intent Security Engine learns what each agent is meant to do from its configuration and its history, and it checks every action the agent takes against that purpose.
An action that does not fit, such as a support agent starting a funds transfer in the middle of a refund conversation, is flagged, masked or blocked inline in under 50 milliseconds, with 98.6% detection accuracy. Because the check is made against the agent’s purpose, it catches an agent that has gone off task even when nothing in the request looks like a known attack.
Lasso’s detection is built for the size of agentic inputs. It holds the same accuracy and latency as the context grows toward a million tokens, so an injection hidden deep in a long input is found before the agent acts on it.
The Lasso platform
The platform covers the full path from finding an agent to protecting it in production.
- Discovery finds every AI agent, model and component in use across SaaS tools, code, cloud and agent builder platforms.
- AI-BOM gives each agent a bill of materials listing the models, MCP servers, tools and components it depends on, so a vulnerable component can be traced to every agent that uses it.
- AI-SPM maps each agent’s connections in a visual graph that shows its blast radius, checks configurations against NIST, OWASP and MITRE, and limits use to approved models and components.
- Automated AI Red Teaming tests agents the way an attacker would. Reconnaissance of each agent shapes static, dynamic and high-agency attacks drawn from more than 300,000 payloads, and the tests run in CI/CD on every deployment.
- Runtime Protection and AI Detection and Response enforce policy per agent at the SDK, API or AI Gateway layer and detect prompt injection and other OWASP and MITRE risks. Each alert explains what happened, which agent was targeted and what to do next.
- Rogue Agent Protection uses the Intent Security Engine to detect goal misalignment, intent drift and behavioral anomalies, and it stops an agent’s action when it moves outside its purpose.
Customer success stories
A European insurance group runs Lasso on the production agents that process its internal claims, where prompt injection attempts arrive embedded inside the claims documents themselves. Lasso blocks them before the agents act. The company has since moved the application to production and is adding hundreds of new agents.
A global hospitality company uses Lasso to secure the agents it built for employee-facing logistics workflows. Intent-based guardrails enforce each agent’s behavior in real time, and the company’s internal MCP-based use cases were secured without code changes.
A global automotive manufacturer runs Lasso’s runtime protection across its customer-facing systems to stop jailbreak attempts and sensitive data exposure. A red teaming engagement on a separate chatbot found more than 15 vulnerabilities that had gone undetected, and all of them were fixed before they could be exploited.
Lasso Security is part of the NetApp Excellerator program. If you’d like to learn more about Lasso, book a demo today.


Comment
0Upholding journalistic discourse. No registration required.
No comments yet. Share your perspective!