A new product watches what agents actually do, not just what they report doing.
SecurityWeek reports that Sysdig has launched Sysdig AI Defense, a runtime security and governance product built on the company's existing visibility stack. It monitors every action an agent takes from start to finish, drawing on kernel-level telemetry to catch activity an agent's own logs might miss, and builds a live inventory of agents, models, MCP servers, skills and plugins, with every agent traced back to an accountable person.
Policy templates map to the OWASP Top 10 for LLM Applications, starting in alert mode before an organization moves them into enforcement, where risky actions can be flagged, sent for human review, or blocked outright. The system logs actions as they happen into audit quality, immutable records showing which policies applied and who signed off. Sysdig's argument, per the report, is that an agent's self-reported account of what it did is not sufficient evidence on its own, and a mismatch between what an agent reports and what its actions show is itself a detection signal.
Why it matters: as enterprises hand agents real access to systems and data, visibility into intent is not enough, security teams need runtime proof of what an agent actually did and who is accountable for it. Treat agent actions like privileged human actions: inventory them, govern them against a known framework, and keep an audit trail that does not rely on the agent's own word.