CrowdStrike plans to acquire Fiddler AI and Dreadnode for about $495 million
CrowdStrike announced plans to acquire two startups focused on AI agent security, Fiddler AI and Dreadnode; The Wall Street Journal reported the deal value at $495 million, with Fiddler accounting for about four-fifths. Fiddler provides a pre-launch risk testing platform for agents and uses a custom model called Centor to block malicious prompts and harmful outputs in production environments, reportedly adding less than 80 milliseconds of latency; Dreadnode discovers vulnerabilities by simulating cyberattacks, covering more than 100 intrusion techniques, and checks the risks of the MCP servers used by agents. CrowdStrike is also adding an AI Kill Switch feature to Falcon Guardian, which can manually shut down misbehaving agents, cut off their network access, and prevent them from restarting.
Why read this
AI score 62
The two acquisitions correspond respectively to pre-launch risk testing and simulated attacks for agents, showing that AI security is being folded into existing endpoint security product lines.
Topics
Body
CrowdStrike Holdings Inc. today disclosed plans to buy Fiddler AI Inc. and Dreadnode Inc., two startups that specialize in securing artificial intelligence agents.
The company announced the deals alongside an update to its Falcon Guardian product.
The Wall Street Journal reported that CrowdStrike will spend $495 million on the acquisitions. Fiddler is believed to account for about four-fifths of that sum. The Palo Alto, California-based startup sells a platform that developers use to test their AI agents for risks before launching them. The software spots issues such as malicious AI-generated code and training data leaks.
Fiddler can try different model and configuration combinations to find the one that minimizes an agent’s risk level. If the agent still generates prompts that breach cybersecurity rules, developers can use the platform to study the reasoning workflow that produced them. Fiddler says that its software makes it easier to identify the specific processing step that went wrong and fix it.
Once an agent is in production, Fiddler uses a series of custom models called Centor to block threats. It filters malicious prompts and blocks harmful output. According to the company, Centor adds less than 80 milliseconds of latency to the requests it reviews.
“As AI security and observability converge, enterprises need to understand their agents’ actions, intervene when they go off course, and explain why,” said Fiddler co-founder and Chief Executive Officer Krishna Gade.
Dreadnode, the other startup that CrowdStrike is acquiring, takes a different approach to securing AI agents. Whereas Fiddler analyzes agents’ prompt processing workflows to identify erroneous steps, Dreadnode launches simulated cyberattacks to reveal vulnerabilities. The company says that its platform uses more than 100 breach tactics to identify weak points.
Dreadnode also looks for vulnerabilities in the MCP servers that an AI agent uses to carry out its work. The platform ranks the issues that it finds by severity, which enables administrators to fix the most urgent ones first.
CrowdStrike plans to integrate Dreadnode’s technology into its Falcon Guardian product. The latter tool can map out the AI agents in a company’s technology environment and ensure that they meet cybersecurity standards. According to CrowdStrike, Falcon Guardian will use vulnerability data from Dreadnode to generate cybersecurity improvement recommendations.
The product is also receiving a second new feature that CrowdStrike revealed in conjunction with the two acquisitions. The AI Kill Switch, as it’s called, enables administrators to manually shut down an AI agent that is exhibiting risky behavior. The feature can also be used to remove the agent’s network access and prevent it from relaunching.
“CrowdStrike introduces a decisive new way to rapidly disrupt dangerous AI activity when risk emerges,” Michael Devins, a director of product marketing at CrowdStrike, explained in a blog post. “This is the first step toward broader enforcement to coordinate containment across the endpoints, identities, cloud environments, SaaS applications, and browsers where agents operate.”
Photo: Robert Hof/SiliconANGLE