CrowdStrike’s AI Agent Security Deals Target Agents That Do Harm With Valid Credentials

CrowdStrike plans to buy two AI agent security companies, Fiddler and Dreadnode, for a combined price of roughly $495 million, according to The Wall Street Journal as reported by TechCrunch. CrowdStrike announced both deals on November 3, 2026 at its Fal.Con EU conference, saying the products will test AI agents before they go live and record what they decide once they are running.

CrowdStrike Is Paying About $495 Million for Two AI Agent Security Companies

TechCrunch’s Ram Iyer reported, citing The Wall Street Journal, that CrowdStrike is paying around $400 million for Fiddler and about $95 million for Dreadnode. CrowdStrike’s own announcements, both written by Michael Sentonas, do not disclose the prices. Iyer reported PitchBook figures showing Fiddler was founded in 2018 and has raised about $99 million, while Dreadnode was founded in 2023 and has raised $14 million. CrowdStrike describes both as intended acquisitions, and its posts say closing depends on conditions being satisfied. Each product is scheduled for general availability in the fourth quarter of CrowdStrike’s fiscal 2027.

CrowdStrike Argues an AI Agent Can Cause Harm Without Being Hacked

The Fiddler announcement builds its case on agents that behave badly inside their permissions. “AI agents don’t have to be compromised to be dangerous,” Sentonas wrote. His example is a refund agent that uses approved tools and valid credentials, makes a wrong call, and approves a refund that breaks company policy. Conventional security controls would log that activity as legitimate. CrowdStrike says the missing piece is a record of each agent’s decisions, actions, changes and tool calls, along with an explanation of why it acted. Sentonas wrote that the ability to trace agent activity is becoming a priority for auditors, regulators, board members and risk managers.

Fiddler Will Trace AI Agent Decisions Inside Falcon Observe

Fiddler will become Falcon Observe within CrowdStrike’s Falcon platform. Before deployment, CrowdStrike says teams can compare prompts, models and configurations to check that an agent performs its task safely and accurately. After deployment, Falcon Observe traces decisions and reasoning paths across the full hierarchy of agents and reconstructs the decision process so a failure can be tied to the step that caused it. CrowdStrike says it covers first-party and third-party agents, instruments them through OpenTelemetry, and supports LangGraph, Amazon Bedrock, AWS Strands and Google ADK. The company plans to join that agent telemetry with its existing endpoint, identity, cloud and runtime data.

Dreadnode Will Attack AI Agents Before They Reach Production

Dreadnode handles the earlier half of the lifecycle. In a second post, Sentonas wrote that security testing built for conventional, deterministic software cannot find weaknesses inside the reasoning loop of a model or agent. Users describe what they want tested in plain language, and Dreadnode runs more than 50 attack techniques against models, model repositories, applications, agents and MCP servers. Those techniques include jailbreaks, prompt injection, multi-turn attacks and attacks aimed at agent tool use. Findings are ranked by severity and mapped to the OWASP Top 10 for LLMs, OWASP Agentic Security ASI01 to ASI10, MITRE ATLAS and the NIST AI RMF. Reviewers can open the exact attacker prompt and the model’s response, reclassify findings, and export a report for a go or no-go decision. Results will feed Falcon Guardian, CrowdStrike’s runtime protection for AI agents, as policy recommendations.

Why This Matters for Engineering Leadership

CrowdStrike’s refund example describes a failure that conventional controls, by CrowdStrike’s account, record as legitimate. An agent that holds valid credentials and calls approved tools produces logs that look clean, so the only way to catch a bad decision is to keep a record of the decision itself. Engineering leaders can start on that before any vendor product ships in Q4 FY27. Instrument agents with OpenTelemetry now, capturing each tool call, its inputs and the reasoning that preceded it, so the trace exists whichever platform you later adopt. Treat adversarial testing as a release gate with a named owner who signs the go or no-go, in the same way a security review gates a service launch. Ask who in your organization can answer an auditor’s question about why a specific agent took a specific action last Tuesday. If the answer is nobody, that gap is your first agent security project.

FAQ