CrowdStrike Pays $495 Million for Two AI Security Startups as Agent Deployments Outrun the Defenders

CrowdStrike has agreed to buy two artificial intelligence security startups for a combined $495 million, its largest push yet into the market for watching over the autonomous software agents that enterprises are now letting loose inside their own networks.

The larger deal values Fiddler AI, a Palo Alto company founded in 2018 that builds software to monitor and explain what a model or an agent is doing and why, at roughly $400 million. The second, Dreadnode, is a 2023-vintage automated security testing platform that probes models and agents for weaknesses before they are switched on, and carries a price of about $95 million. CrowdStrike said the consideration is mostly cash, with the remainder in restricted stock subject to vesting.

"We are well past the testing phase. Enterprises are deploying AI agents at scale," CrowdStrike president Michael Sentonas said. The risk, he argued, is rarely a machine going rogue. "It is not the agent running wild. It is misconfiguration, people not understanding how it works, permissions granted beyond what is needed, so it reaches resources it should never touch."

The two acquisitions split the agent lifecycle in half. Dreadnode covers the left side, scanning models and agents for jailbreaks and other weaknesses before deployment. Fiddler covers the right side, tracing an agent's decisions, the tools it calls and the exchanges between agents once it is live. CrowdStrike plans to package Fiddler's technology as Falcon Observe and to ship both capabilities into its Falcon platform early next year.

Falcon Observe is designed to sit alongside a new emergency stop in the company's Falcon Guardian product, which lets an administrator terminate a misbehaving agent with a single action. Watch it, then kill it: that combination is quickly becoming the standard shape of the agent security pitch across the industry.

The urgency comes from how agents are bought rather than how they are built. Business units can sign up for an agentic tool without telling the security team, and the agent then runs on legitimate corporate credentials. A misconfigured agent can read files it was never meant to see, or burn through compute, with no attacker present at all. When several agents work in sequence, a single instruction can trigger a long chain of actions that is very hard to trace back to its origin.

Security buyers have noticed. Momentum Cyber counted 117 acquisitions in the third quarter alone, putting 2026 on pace for roughly 450 deals against 404 last year, and AI security acquisitions reached 40 in the first nine months, up from 10 in all of 2025. CrowdStrike's shares fell 3.4% to $241.81 on the day, in line with a broader pullback across AI-linked stocks.

What it means: the agent observability layer is consolidating into platform vendors faster than the startups selling it can reach scale on their own. Within two years the interesting question will not be whether an enterprise can see what its agents did, but whether those logs survive an incident review, and whether the stop button is wired to something that actually stops.