Sysdig has launched AI Defense, a security product designed to monitor AI agents as they act and stop risky behaviour before it reaches production systems. The key idea is to judge an agent by what it actually does, not just what it says it did.
Watch Desk analysis
What happened
SecurityWeek reports that AI Defense uses runtime telemetry to build an inventory of agents, models, tools and related components. Sysdig says organisations can trace each agent to an accountable person, inspect its actions in a live action graph, and detect, hold for human review or block risky actions before they complete.
Policies can start in alert mode and move to enforcement when an organisation is ready. Sysdig says the records capture which policies were applied, when, and who signed off. The company also says its policy templates map to frameworks including the OWASP Top 10 for LLM Applications.
Why it matters
As agents move from answering questions to using tools and changing systems, a security team needs more than a chat transcript to understand what happened. Runtime monitoring could give defenders a view of the underlying activity, while the option to pause or block actions offers a practical control point.
Those are Sysdig’s descriptions of its product, not independent evidence of how accurately it detects risky behaviour or how well it performs under load. Still, the safeguards it proposes are concrete: monitor actions in real time, connect agents to accountable people, and let teams choose when a policy moves from warning to enforcement.
Our read
This addresses a real operational gap: an agent’s account of its own actions is not the same as an audit trail. Sysdig’s emphasis on runtime evidence is sensible, and starting policies in alert mode gives teams a way to observe before they block.
The useful test is whether teams can tune those controls without drowning in alerts or stopping legitimate work. Security products are very good at promising visibility; the hard part is making that visibility useful at 3am.
What to watch
- Whether Sysdig publishes technical detail on detection and blocking latency.
- How teams distinguish genuinely risky actions from legitimate agent work.
- Whether customers can show that the action records help investigations and audits.
Discussion spark: Should AI agents’ risky actions be blocked automatically once policies are set, or should a person approve every intervention?
Sources and evidence
- Source update (4 November 2026, 14:34 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.