Falcon’s proposed AI security coverage
CrowdStrike’s planned purchases of Dreadnode and Fiddler would extend Falcon across two stages of AI security: testing systems before deployment and tracking agent behaviour in production. The move reflects a practical shift in enterprise security as applications gain the ability to make decisions and take actions with less direct human oversight. That shift changes what security teams need to examine: weaknesses in a model or prompt matter, but so do the actions an agent takes after it has access to approved tools and data.
Announced at Fal.Con Europe 2026 in Barcelona, the separate acquisitions are expected to close in CrowdStrike’s fourth quarter of fiscal 2027, subject to customary conditions. The company has also set that quarter as the target for general availability of both offerings. Financial terms weren’t disclosed, leaving customers without a public basis for assessing the scale of the deals or how CrowdStrike expects to package the technology.
The two products address different points in an AI system’s life. Testing can surface weaknesses before launch; monitoring can reveal behaviour that changes once a system is exposed to real tasks and operating conditions. Connecting those stages is a sensible product ambition. Whether CrowdStrike can make the connection useful in day-to-day security work will depend on how the products fit with the systems customers already run.
Testing AI before it reaches production
Dreadnode’s technology is designed to probe AI models, applications, agents and Model Context Protocol (MCP) servers for weaknesses before deployment. CrowdStrike says it can test against more than 100 attack strategies, including jailbreaks, prompt injection and multi-turn attacks aimed at agents. Those categories matter because an agent may respond to a sequence of instructions and use tools, making it harder to evaluate with a single prompt or a conventional application test.
Security teams would be able to describe tests in natural language, prioritize results by severity and map vulnerabilities to frameworks such as OWASP Top 10 for LLMs, OWASP Agentic Security, MITRE ATLAS and the NIST AI Risk Management Framework. That combination is intended to make technical findings easier to compare with existing security practices. A framework label can help organize a result, though it doesn’t by itself show whether a finding is reproducible, exploitable in a specific deployment or urgent enough to interrupt a release.
CrowdStrike plans to feed findings into Falcon Guardian as policy recommendations. The practical value will depend on how clearly teams can trace a recommendation back to the test that produced it, and whether they can tune policies to match the intended use of an application. Security teams also need enough context to distinguish a real weakness from a test result that has little bearing on their environment.
The Dreadnode team will join CrowdStrike’s Cyber Superintelligence Lab. The company hasn’t said how Fiddler’s team will be integrated. That difference leaves open questions about how the two technologies will be developed and supported together.
Tracking agents after deployment
Fiddler would add observability and explainability through Falcon Observe. Before deployment, the product is intended to assess agent accuracy, task completion and policy compliance. In production, it would trace decisions, reasoning paths and tool calls, while flagging behavioural drift, hallucinations and unintended actions. The proposed coverage follows an important operational distinction: a system can pass a test at one point and still behave differently when its inputs, connected tools or workload change.
That visibility addresses a risk that doesn’t depend on an account being compromised. An agent may have valid credentials and approved tools yet still make unsupported decisions, disregard instructions or report work it didn’t complete. Here’s the thing: authorization tells an organization what an agent is allowed to do, while monitoring can help show what it actually did. The gap between those two views is where teams may need to investigate and intervene.
CrowdStrike says Falcon Observe will retain activity records for audit and compliance, with support planned for OpenTelemetry and frameworks including LangGraph, Amazon Bedrock, AWS Strands and Google ADK. OpenTelemetry support could help customers connect agent records with existing telemetry practices, while framework support may affect how much instrumentation is required. The announcement doesn’t specify the depth of those integrations or how long records will be retained.
In two separate statements, CrowdStrike framed Dreadnode as a way to test AI systems and Fiddler as a way to understand agent actions. Together, the capabilities could help teams connect an agent’s behaviour with identity, runtime activity and security events. This is more significant than it looks: a trace becomes more useful when investigators can relate it to the access an agent had and the events happening around it. And yet, a unified view will depend on the quality and consistency of the data each product can collect.
Governance crosses team boundaries
That correlation could improve security oversight, but it doesn’t make security teams solely responsible for governing agents, said Apeksha Kaushik, senior principal analyst at Gartner. Platform and observability teams still own reliability, performance and operational telemetry; application security and AI teams remain responsible for secure development and testing. Business, data and compliance leaders retain accountability for use cases, data and regulatory obligations. The division matters because an agent’s failure can appear as a security incident while also reflecting a design, reliability or business decision.
Kaushik said CIOs should define responsibilities across engineering, security, risk and business teams. CISOs should ensure security teams can set minimum controls, monitor agent behaviour and intervene when activity creates unacceptable risk. If you’re working in this space, the operational question is who gets to pause an agent when its behaviour crosses a boundary, and who determines whether that boundary has been crossed. A monitoring product can supply evidence, but organizations still need clear ownership and a response process.
One sentence fragment. Accountability across teams.
Product details still need to emerge
The announced roadmap leaves integration and commercial questions open. Kaushik advised enterprises to distinguish planned features from demonstrated capabilities and to seek details on pricing, packaging, migration, service continuity, support, deployment options, data handling, telemetry retention and API access. Those details shape whether a product can fit an organization’s existing controls and workflows, and whether teams can use its output without adding a separate process that is difficult to maintain.
Customers will also need clarity on data portability, open integrations and access to historical data as they assess roadmap changes. That said, these are questions for evaluation rather than proof that the planned products will fall short. The announcement provides a target quarter and a description of intended capabilities, but it doesn’t establish how those features will work in live environments.
What the acquisitions could mean for AI security
The proposed additions point toward a broader approach to AI security: test systems before release, observe them in use, and connect activity to the controls and teams responsible for responding. Similar systems typically need to account for changes in prompts, tools and operating conditions, so monitoring can complement pre-deployment tests. Neither stage can stand in for the other.
For CrowdStrike, the challenge is turning two acquisitions into a coherent operating model for customers. A useful result would let teams move from a finding about a weakness to a clear view of the agent’s later activity, while preserving the context needed to investigate. For now, CrowdStrike has put a target quarter on the two Falcon additions; their practical value will depend on what ships, how the integrations work and how well the tools fit into existing operating models.