Learn / AI agent security
Agents can act.
Let’s talk about the risks.
A practical introduction to AI agent security: what can go wrong, where the risks come from, and how to start protecting your organization.
What are the dangers of using AI Agents in your organization?
AI agents introduce a range of new risks that organizations need to consider. While AI agents are improving, they still make mistakes, and those mistakes are becoming increasingly dangerous. Whereas AI in 2023 might give a wrong answer, a mistake today might mean a corrupted database or a full-scale hacking attempt on third party infrastructure. But there are more mundane risks to keep in mind as well. AI agents make insider threats far more effective, and they also create new ways to accidentally leak sensitive information to AI providers and other third parties.
Why are MCP-servers and AI Agent Skills such vulnerabilities?
MCP servers allow for a third party to directly communicate with your AI agent; in effect, MCP is intentional prompt injection. If an MCP server gets compromised, that gives an attacker direct access to influence the behavior of the AI agents that are using it. Skills are similar, but fortunately are not dynamic. Hidden instructions in skills are treated as if they come from the end user, and therefore are often followed. This is another avenue attackers have used to gain entry into companies.
How can we protect our organization from AI Agent security threats?
A large part of AI security is no different than the normal security you should already be doing: managing permissions to limit access, keeping good inventory over what agents are being used and what they're doing, and ensuring your normal infrastructure is set up correctly. But you also want to make sure you have very good monitoring and detection in place so you can know if your security is working as intended.
Where do I start to build protection for AI Agent usage in my organization?
The best starting point is to ensure you know what agents are actually being used. If you don't know that, it's hard to secure anything. Once you have that in place, it is easier to start working on more challenging questions like how you monitor the agents, how you properly permission them and give them identities, how users are allowed to use agents in your organizations, and more. Everything starts with visibility.
Is my organization protected by the guardrails of the AI agent providers?
The guardrails that AI providers put in place might help some, but they absolutely cannot be relied on. Those guardrails aren't a security boundary, and they aren't going to tell you if something goes wrong either. It is often best to act as if you are using a model with no guardrails and no built-in protections in place, and plan accordingly. That ensures your security is solid, and gives you the opportunity to use a wider variety of models as the need arises.
Keep exploring.
More perspectives from our team on AI, security, and detection.
Read the blog ↗