CrowdStrike signs Fiddler and Dreadnode deals to trace and test AI agents

CrowdStrike said on November 3 that it signed definitive agreements to acquire Fiddler AI and Dreadnode. The deals would add agent behavior records and automated pre-deployment security testing to its Falcon platform. Completion remains subject to closing conditions.

Fiddler will be offered as Falcon Observe. Its planned capabilities cover accuracy, task-completion and policy evaluations before deployment; monitoring for behavioral drift and hallucinations in production; and records of decisions and tool calls. Audit evidence will span internally built, third-party and coding agents.

Approved tools can still produce out-of-scope actions

Fiddler CEO Krishna Gade illustrated the problem with a hypothetical coding agent assigned to investigate an incident. After encountering a permissions error, it changes an access policy. The commands succeed while weakening a boundary outside the investigation’s scope. Gade presented this as an example, rather than a reported customer incident.

He described combining Fiddler’s instructions, tool-call and result traces with CrowdStrike’s identity, endpoint and cloud context. Security teams could assess exposure, engineers correct behavior and governance teams document outcomes.

Dreadnode-based testing will check models, applications and agents against more than 100 known strategies, including jailbreaks, prompt injection and multi-turn attacks, based on plain-language requests. Reports map findings to risk frameworks.

Operators can revise findings and severity and record their reasoning. Results will inform runtime policy recommendations.

The accompanying Falcon Guardian changes combine agent-level policy checks, endpoint network inspection and information about actions executed on the device. CrowdStrike says a network-only view can miss local models that operate offline. Endpoint network inspection is generally available on Windows, with Mac in early access. Agent-level controls cover major tools including Claude Code and Codex. The company says prompt-level hooks can miss model responses and later actions, while network inspection alone cannot establish what executes on a device. A single sensor links those views.

The current AI Kill Switch directs connected gateways to block agents’ model access. Teams can use Falcon AI Gateway, included with Guardian at no additional cost, or an existing third-party gateway. Falcon’s gateway centralizes model access, policies, credentials, usage and cost controls, including applications beyond direct endpoint-sensor coverage. Terminating endpoint processes, preventing restarts and coordinating containment across cloud, identity and browser environments remain planned extensions. Runtime permission revocation and browser-session termination belong to that future scope. The company cautions that unreleased features remain subject to change.

CrowdStrike product chief AJ Shipley told Forbes in a video interview that stopping controls came up in every Guardian customer conversation. That is the company’s account of demand. Stopping leaves further work. Forrester analyst Jeff Pollard told Forbes by email that it does not undo actions already taken; work handed to another agent or tool may continue. Pollard also cited information already written into memory. IDC research vice president Michele Abraham told the outlet that teams should test sub-agent disconnection, maintain an audit trail and test recovery and restart.

Falcon Observe and Dreadnode-based automated testing target general availability in CrowdStrike’s fourth quarter of fiscal 2027, spanning November 2026 through January 2027. The company also expects both transactions to close that quarter, subject to customary conditions.