CrowdStrike acquires Fiddler and Dreadnode: security for AI agents

Two acquisitions, one same idea: monitor AI before and after it acts

CrowdStrike has announced at Fal.Con Europe 2026, held in Barcelona, the acquisition of two artificial intelligence security companies: Fiddler AI and Dreadnode. Both transactions will close as independent businesses and their technology will end up integrated into the Falcon platform. The underlying message is clear: security is no longer content with protecting the infrastructure where AI runs, but wants to get inside the model itself and the agent's behavior.

The company expects to close both transactions in its fourth quarter of fiscal year 2027, subject to customary conditions, and for the resulting capabilities to be generally available in that same quarter. That is, it is not something an SMB can contract tomorrow, but it is a signal of where the market is heading.

Dreadnode: testing AI before putting it to work

Dreadnode provides automated security testing to identify vulnerabilities in models, applications, and agents before their deployment. Its team will join CrowdStrike's Cyber Superintelligence Lab.

According to the information provided, its technology will allow security teams to test models, agent reasoning loops, and Model Context Protocol (MCP) servers against more than 100 AI attack strategies. Among them are mentioned jailbreaks, prompt injection, multi-turn attacks, and those specifically targeting agents.

There is a detail that seems especially useful to us for small teams: tests can be described in natural language, findings classified by severity, and vulnerabilities mapped to frameworks such as OWASP Top 10 for LLM, OWASP Agentic Security, MITRE ATLAS, and the NIST AI Risk Management Framework. Findings will be incorporated into Falcon Guardian as policy recommendations.

Fiddler: understanding why the agent did what it did

The second acquisition targets a risk that many companies have not yet internalized: an agent can have valid credentials and authorized tools and, even so, make unauthorized decisions, ignore instructions, or claim work it has not completed. No one needs to compromise it for it to generate a problem.

Fiddler will add agent observability and explainability through Falcon Observe. The idea is to evaluate agent accuracy, task completion, and policy compliance before deployment, and then trace decisions, reasoning chains, and tool calls already in production. Behavior deviations, hallucinations, and unintended actions would also be detected, maintaining records for compliance and audit purposes.

The planned offering will be compatible with OpenTelemetry and with frameworks such as LangGraph, Amazon Bedrock, AWS Strands, and Google ADK.

Governance is not bought, it is shared

Analyst Apeksha Kaushik, from Gartner, provides the most sensible reflection in the announcement: incorporating observability and testing into a security platform helps correlate agent behavior with identity, runtime activity, and incidents, but does not make the security team the sole party responsible for AI governance.

Platform and observability teams remain responsible for reliability, performance, and operational telemetry; application and AI security teams, for secure development and testing; and business, data, and compliance, for use cases, data, and regulatory obligations. Kaushik recommends that CIOs define a governance model that distributes responsibilities among engineering, security, risk, and business, and that CISOs ensure their teams have the authority to set minimum controls, monitor agent behavior, and intervene when risk is unacceptable.

What this means for an SMB

If your company is starting to set up agents or AI assistants on real processes, this move confirms a trend we had already been seeing: the market will sell AI security as one more layer of the stack, not as a separate project. And that, in the short term, has three practical consequences.

• Do not wait for the product to have criteria. Although general availability arrives in the fourth quarter of fiscal year 2027, the frameworks they cite (OWASP LLM, OWASP Agentic Security, MITRE ATLAS, NIST AI RMF) are public and free. They are a good checklist for reviewing which agents you have and what they can touch.
• Distinguish between "compromised" and "gone astray." The risk of the agent that acts badly without anyone having attacked it is real and often goes unnoticed. If an agent can send emails, modify records, or launch requests to internal systems, it needs decision traceability, not just access logs.
• Prepare the agent's identity. Its own credentials, least privilege, and separation of duties. It is the same hygiene you would apply to an employee or a service, and it fits with what we already discussed in our reflection on AI in the company.

Our recommendation from ForgeNEX

For an SMB in Seville or of any medium size, the temptation is to buy the tool and delegate the problem. It does not work that way. Before looking at licenses, it is advisable to do three things:

1. Inventory the AI agents and assistants already in use, even if they are informal pilots in spreadsheets or HR tools.
2. Define what decisions each one can make, with what data, and with what limits. Document it, even if only in a table.
3. Agree who responds if an agent does something improper: business, IT, or the provider. Without that distribution, observability only generates reports that no one uses.

The acquisition of Fiddler and Dreadnode is good market news, but it does not replace governance. As we commented when discussing AI agents and the IT channel's business model, the real opportunity lies in supporting adoption with criteria, not in reselling one more console.

Source: Computerworld. Analysis and adaptation: ForgeNEX.