THE CRUNCH
Sysdig has announced AI Defense, a product that applies the company's runtime visibility technology to agentic AI, monitoring what agents actually do and blocking risky actions before they complete. The pitch is that AI has shifted from answering questions to taking actions: agents now ship code, change infrastructure and reach into production systems faster than any team can review by hand. Sysdig founder and CTO Loris Degioanni argues that an agent's own report of what it did is useful context but not evidence, so the product watches behaviour directly rather than trusting the agent's account.
The technical challenge, as Sysdig describes it, is that agents can change rapidly and invisibly. Underlying every logged agent action there may be five or more processes visible only to the operating system kernel, and these could be quietly altering what the agent does. The company's warning is blunt: the agent deployed in the morning is not always the one running in the afternoon, and only the kernel shows what actually ran. Blocking, where needed, must happen in line and in real time, before the action completes.
The product builds a live inventory of agents, harnesses, models, MCP servers, skills and plugins from runtime telemetry, and traces each agent to an accountable human. Risky actions can be detected, held for human review or blocked. Policy templates map to frameworks such as the OWASP Top 10 for LLM Applications, and each policy starts in alert mode, moving to enforcement only when the organisation is ready.
Every action is recorded as it happens in a live action graph, pairing what the agent said it was doing with what it actually did, what it could reach and whose authority it acted under. The result, Sysdig claims, is audit-quality immutable records showing which policies were applied, when, and who signed off.
The broader problem is scale. An IBM Institute for Business Value report from earlier this year found enterprises expect to deploy an average of 1,661 AI agents by 2027, a 38% increase on today, producing volumes of autonomous decisions that manual governance cannot realistically supervise.