Without adequate governance, AI agents risk autonomously, invisibly, and rapidly transitioning from an asset to a liability for businesses. Sysdig has announced Sysdig AI Defense, a product that brings runtime security and governance to these agents.
- Sysdig launches AI Defense to secure and govern AI agents at runtime.
- The product monitors every agent action in real time and can block risky actions before they complete.
- It maintains a live inventory of agents, each linked to a human owner, with policies that start in alert mode.
- An immutable audit log compares what agents report with what they actually do.
For Loris Degioanni, founder and Chief Technology Officer (CTO) of Sysdig, the need is clear. “AI is moving from response to action. First there were chat machines. Today, agentic AI does things on our behalf,” he explains.
The problem lies in these agents’ ability to go further than requested, faster than can be controlled. Sysdig highlights the need to continuously monitor what they are doing.
Tracking every action, all the way through
According to Sysdig, AI Defense observes and monitors every action of an agent, from its starting point to its completion. The task is far from simple. An agent can quickly acquire new capabilities while seeking to accomplish its mission. Each action can be logged, but behind it can hide at least five processes, visible only to the operating system kernel, that discreetly modify the agent’s operation.
“The agent deployed in the morning is not always the one running in the afternoon,” warns Sysdig, adding that “only the kernel shows what was actually executed.” If security requires blocking an action, the response must be immediate. “We need solutions that work at runtime, very quickly, and can trigger a response,” emphasizes Loris Degioanni. “Any blocking response must occur in real-time, inline, while the action is unfolding and before it is completed.”
For the executive, the distinction between what the agent declares and what it does is central. “An agent can misreport a tool call, but it cannot rewrite its interactions with the system or the outside world,” he states. “The narrative an agent gives of its actions is useful context, but it is not proof. When what it reports and what its actions show do not match, you have a detection.”
Inventory, policies, and traceability
Sysdig claims its product offers several distinct functions. It establishes a live inventory of agents, their harnesses, models, MCP servers, skills, and extensions, built from runtime telemetry. Each agent can be linked to an identified human owner.
Agentic actions deemed risky can be detected, subjected to human validation, or blocked before being completed. Predefined policy templates can incorporate rules associated with frameworks like the OWASP Top 10 for LLM Applications. Each policy starts in alert mode, then switches to enforcement mode when the organization is ready.
The agent’s actions and those of AI Defense are recorded as they happen. A live action graph presents each step in order, comparing what the agent says it is doing with what it is actually doing, what it can access, and under whose authority it is acting. The result, according to the company, is an immutable audit log showing which policies were applied, when, and who approved them.
The topic fits into a broader context. The IBM Institute for Business Value, in a study dedicated to agentic AI governance, notes that “as AI agents move from pilot projects to production, systems make autonomous decisions at volumes that no human governance can reasonably oversee.” According to this study, “by 2027, companies plan to deploy an average of 1,661 AI agents, an increase of 38% from today.”
