CrowdStrike buys Fiddler to bring agent behaviour under the security umbrella

The deal

CrowdStrike has signed a definitive agreement to acquire Fiddler AI, the company says, with the aim of covering both the environments agents run in and the behaviour of the agents themselves. CrowdStrike already protects agents and the environments they operate in, according to Fiddler, while Fiddler traces instructions, tool calls and results and evaluates whether an agent stayed within its task. m:…9029

The combination is meant to help enterprises decide which agents are ready to deploy, what they are allowed to do, and when they can be trusted with more responsibility, Fiddler says. That framing matters because the failure modes are not only attacks: Fiddler notes that agent failures can happen without an attacker, and that attackers can also exploit the same weaknesses. m:…9029

Fiddler was started eight years ago on the belief that businesses would need to understand the AI they depended on, and its work since has covered explainability, model monitoring, evaluations, agentic observability and runtime guardrails. Its Centor evaluator models provide a foundation for this work, running at a fraction of the cost of large language models to assess faithfulness, sensitive information and safety, the company says. m:…9029

Why buyers want this now

The same week brought two more signs that agent security is becoming a product category of its own. Arthur AI released Deckard, an open-source scanner that runs routine scans across endpoints, cloud providers and managed agent platforms to keep an inventory of every agent running in an organisation, with data flowing into existing security information systems so analysts can spot rogue agents and act quickly. m:…8799

Deckard ships with signatures for osquery, deploys through fleet and device management platforms such as Fleet and Jamf, analyses network logs to detect agents calling language models and other agentic resources, and covers cloud products from AWS Bedrock to Gemini Enterprise as well as managed platforms such as the OpenAI Agents API and Claude Managed Agents, Arthur says. The pitch is blunt: organisations cannot secure agents they cannot see. m:…8799

On the offensive side, RunSybil says it deploys an average of 2,700 coordinated agents per engagement in a layered hierarchy, sharing context in real time so low-priority findings across an application can be connected into critical, chainable exploits. Its reports include CWE identifiers, CVSS scores and full request and response evidence, delivered inside a SOC 2 Type II environment with per-customer isolation, the company says. m:…8429

What changes for security teams

A SANS Spotlight paper by Chris Dale makes the case for continuous validation over point-in-time testing: traditional penetration tests reflect a moment in time, while applications, identities, infrastructure and attacker techniques keep changing. Autonomous systems, in that paper's account, combine deterministic techniques, rule-based reasoning, machine learning and language models, operating within defined authorisation and safety boundaries while security professionals set objectives and oversight. m:…6160

The paper's workflow runs from discovery through exploit validation and chained attack paths to demonstrated business impact, which then informs remediation priorities, and repeat assessments verify that fixes held. That is the same logic as the Fiddler deal: the question is no longer only whether an agent was breached, but whether its behaviour stayed within bounds, and whether the evidence for that claim can be checked after the fact. m:…9029 m:…6160

For buyers, the practical test is integration. CrowdStrike brings context about identity, endpoint activity and cloud resources, Fiddler says, and Fiddler connects that protection to whether agents themselves are behaving safely. Enterprises evaluating the combined offering should ask what happens to existing Fiddler customers, how Centor evaluations map to their own deployment gates, and whether guardrails run inline or after the fact. m:…9029