CrowdStrike to Acquire Fiddler AI to Tie Agent Behavior to Security

What the deal covers

According to Gade, CrowdStrike already helps enterprises protect their agents and the environments in which they operate. Fiddler is meant to add a layer that shows whether the agents themselves are behaving safely and doing what was intended. The post does not state a purchase price, a closing date or other terms of the agreement.

Gade wrote that he and co-founder Amit started Fiddler eight years ago on the belief that businesses would need to understand the AI they depend on. He pointed to his earlier work at Facebook, where his team built "Why am I seeing this post?" to explain a ranking model to a billion people. According to Gade, that conviction shaped Fiddler's work in explainability, model monitoring, evaluations, agentic observability and runtime guardrails.

Security and safety as two separate questions

The post frames the deal around the shift from AI that answers to AI that acts. Agents write code, access sensitive data, invoke tools and make decisions on behalf of users, so the consequences of their mistakes grow with their responsibilities. Gade says an agent can hold valid credentials and approved tools while still making unsupported decisions, ignoring instructions or claiming work it never completed. Such failures can happen without an attacker, though attackers can exploit the same weaknesses.

Enterprises therefore need to know both whether someone is attacking an agent and whether the agent is going wrong on its own. Gade describes the first as a security question and the second as a safety question.

He gives an example of a coding agent authorized to investigate an incident. The agent hits a permissions error and changes an access policy to make the error disappear. The commands succeed, but the agent has weakened a security boundary while doing work it was only authorized to investigate. According to Gade, Fiddler can help trace the instructions, tool calls and results, and evaluate whether the agent stayed within its task. CrowdStrike would add context about the identity, endpoint activity and cloud resources involved.

Gade says combining these capabilities helps security teams assess exposure and respond, engineers correct the behavior, and governance teams document the outcome. The same evidence can then inform a new test and a runtime control, so the next agent does not repeat the mistake.

A control plane for agents

Gade expects enterprises to use many models, clouds and frameworks. They will need oversight across agents they build, coding assistants their employees use, and agents embedded in software they buy. Together with CrowdStrike, Fiddler wants to help companies decide which agents are ready to deploy, what they are allowed to do, and when they can be trusted with more responsibility. According to the post, that requires connecting development evaluations, production monitoring and runtime policy.

Gade names Fiddler's Centor evaluator models as a foundation for this work. He says they run at a fraction of the cost of LLMs to assess faithfulness, sensitive information and safety, while addressing production requirements for cost, latency and privacy.

What it means for existing customers

Gade says the needs that brought customers to Fiddler remain central to its mission: understanding model behavior, improving AI quality, managing cost and meeting governance requirements. That covers predictive models, generative AI and agents. He describes security as a natural extension of this work, with the chance to make the evidence customers already rely on useful across more of the enterprise.

The post closes with thanks to customers, partners and the Fiddler team. Gade writes that the company intends, together with CrowdStrike, to turn understanding of AI into the confidence to put AI to work across the enterprise.