Sysdig's AI Defense Tackles Rogue Agents

Agentic AI is shifting from answering questions to taking action—shipping code, altering infrastructure, and touching production systems faster than any team can manually review. Sysdig has launched AI Defense, a product that applies runtime security and governance to these autonomous agents, aiming to catch and stop risky behavior as it happens.

From Chat Machines to Autonomous Actors

The core problem, according to Sysdig, is that AI agents now do things for us, but they can do more than instructed, faster than we can control, and without a moral compass. As agents gain autonomy and reasoning capabilities to maximize economic benefit, the complexity increases the chance they drift into error. Sysdig warns that unattended, the risk of agents going rogue will only grow.

“AI has moved from answering to acting, and agents now ship code, change infrastructure, and reach into production systems at a speed no team can review by hand,” the company states. The need for governance is clear: without it, agentic AI could morph from a business boon into a business disaster, rapidly and invisibly.

Kernel-Level Visibility for Agent Actions

Sysdig AI Defense monitors every action an agent takes, from its origin to its completion. But that's no simple task. Agents can assume new powers on the fly as they pursue their tasks. While an agent may log its actions, underneath each action there may be five or more processes visible only to the operating system kernel—processes that could be silently altering the agent's function.

“The agent deployed in the morning isn’t always the one running in the afternoon,” Sysdig warns, adding that “only the kernel shows what actually ran.” This kernel-level telemetry is the foundation of the product's visibility.

Real-Time Blocking, Not Just Logging

If an agent's action needs to be stopped for safety, it must be blocked immediately. Loris Degioanni, Sysdig founder and CTO, explains: “AI is moving from answering to acting. It used to be chat machines. Now agentic AI does stuff for us.” He adds, “There is a need for solutions that operate at runtime, can operate extremely quickly, and can include response. For example, any blocking response must happen in real time, in line, as the action happens and before the action completes.”

“An agent can misreport a tool call, but it can’t rewrite its interactions with the system or the external world. An agent’s own account of what it did is useful context, but it’s not evidence. When what an agent reports and what its actions show disagree, you have a detection. Watching software in real time and stopping it when it steps out of line is the core of Sysdig’s runtime foundation, and Sysdig AI Defense brings that same truth to the moment an agent acts.”

— Loris Degioanni, founder and CTO at Sysdig

What the Platform Delivers

Sysdig AI Defense offers a live inventory of agents, harnesses, models, MCP servers, skills, and plugins, all built from runtime telemetry. Every agent can be traced to an accountable human. The platform detects risky actions, holds them for human review, or blocks them before completion. Policy templates can be predefined with rules mapped to frameworks like the OWASP Top 10 for LLM Applications. Each policy begins in alert mode and can be switched to enforcement mode when the organization is ready.

Actions taken by both the agent and AI Defense are recorded as they occur. A live action graph displays each step in order, pairing what the agent said it was doing with what it actually did, what it could reach, and whose authority it acted under. The outcome is audit-quality immutable records showing which policies were applied, when, and who signed off.

IBM Study Highlights Governance Gap

The need for such controls is underscored by research from IBM's Institute for Business Value, which earlier this year reported on the state of agentic governance. The study notes that as AI agents move from pilots to production, systems make autonomous decisions at volumes no human governance can realistically supervise. It also warns the problem is worsening rapidly: by 2027, enterprises expect to deploy an average of 1,661 AI agents, a 38% increase from today, forcing tech leaders to manage hundreds of thousands of autonomous decisions daily.

• 1,661 AI agents: average expected deployment per enterprise by 2027
• 38% increase: expected growth from today's agent count
• OWASP Top 10 for LLM Applications: framework used for policy templates
• Five or more underlying processes: potentially hidden kernel-level activities per agent action

Accountability and Audit Trails

Sysdig emphasizes that AI Defense provides a clear chain of accountability. Each agent can be linked to a responsible human, and every action is logged with the policies applied and any approvals. The live action graph helps security teams understand not just what an agent did, but also the context and authority behind each step. This addresses a critical gap: agents may misreport their actions, but their actual interactions with systems and the external world are harder to falsify.

Why It Matters for Enterprises

As agentic AI adoption accelerates, the ability to monitor and control autonomous agents becomes a business imperative. Sysdig's approach suggests that runtime security—operating at the kernel level and able to block actions in real time—could be essential for safely deploying agents at scale. Organizations evaluating agentic AI should consider whether their governance tools can keep pace with the speed and complexity of machine decisions, and whether they have the audit trails needed to satisfy compliance and accountability requirements. Without such safeguards, the very autonomy that promises efficiency could introduce unmanageable risk.

Reporting based on original coverage from SecurityWeek. Read the original article here.