LE PERSPECTIVE / THE STORY

CrowdStrike is buying two startups to answer the question every company rolling out AI agents now faces: can you trust what they do? The cybersecurity company announced deals for Dreadnode and Fiddler AI on Nov 3, at its Fal.Con Europe event in Barcelona.

One tests AI before it goes live. The other watches it once it's working. Together they cover the whole life of an AI agent.

Agents can go wrong without being hacked
CrowdStrike's own framing is blunt: an agent with valid credentials and approved tools can still ignore instructions, make calls nobody approved or claim work it never did. The risk is no longer just attackers getting in. It's the AI you already let in.

Dreadnode: break it before launch

Traditional security testing assumes software behaves the same way every time. AI doesn't. Models can be jailbroken, and agents can be steered by hidden instructions buried in the content they read, a trick known as prompt injection.

Dreadnode attacks AI systems before real attackers do. Teams describe what to test in plain English, and it runs more than 100 known attack strategies against models, agents and the tools they connect to. Findings are ranked by severity and mapped to standard frameworks like OWASP and NIST, so teams can make a clear go or no-go call.

Fiddler: a flight recorder for agents

Fiddler covers what happens after launch. It traces each decision an agent makes, the reasoning behind it and the tools it used, and flags drift, hallucinations and actions nobody asked for. CEO George Kurtz described it as a flight recorder for every agent.

That matters the moment a board or regulator asks what an agent did. Kurtz's point is that blaming the model won't be an acceptable answer. CrowdStrike will sell Fiddler's technology as Falcon Observe.

Why it matters

Security companies have spent decades keeping attackers out. AI agents flip the problem: they're inside by design, with real access to real systems. The new job is proving they behave.

With these two deals, CrowdStrike can test an agent before launch, protect it while it runs and explain what it did afterward, all on one platform. For companies stuck in AI pilots, that kind of proof is often what decides whether agents get real work at all.

The bigger signal: AI trust is turning into its own security market, and the biggest players intend to own it early.

Get stories like this in your inbox every week. Subscribe to The Week in Perspective.