CrowdStrike to buy Fiddler AI and Dreadnode, adds AI agent kill switch

CrowdStrike on Tuesday announced agreements to acquire two AI security startups, Fiddler AI and Dreadnode, and launched an AI Kill Switch designed to stop rogue AI agents. The deals are worth a combined $495 million, around $400 million for Fiddler and $95 million for Dreadnode, according to The Wall Street Journal as cited by Stocktwits; CrowdStrike itself did not disclose the acquisition values.

The two purchases target different stages of an AI agent’s life. Dreadnode attacks agents and models before they ship to find weak spots; Fiddler watches and explains what agents do once they run. The Kill Switch covers the moment an agent has already gone wrong. For developers building agents that act on company systems, the announcement is a signal of how large security vendors now expect those agents to be tested, logged and, when needed, shut off.

What CrowdStrike announced and what the deals cost

CrowdStrike said both transactions will be funded primarily with cash, with a portion paid in stock, and are expected to close in its fiscal fourth quarter of 2027. The new product capabilities built from the two companies are scheduled for general availability in that same quarter. The sources give the timing only in CrowdStrike’s fiscal calendar, not as calendar dates.

Item	Detail	Status
Fiddler AI	About $400M (reported by WSJ)	Expected to close fiscal Q4 2027
Dreadnode	About $95M (reported by WSJ)	Expected to close fiscal Q4 2027
Falcon Observe (from Fiddler)	Agent monitoring and explainability	General availability planned fiscal Q4 2027
Automated AI security testing (from Dreadnode)	Pre-deployment attack testing	General availability expected fiscal Q4 2027
AI Kill Switch	Stops agents via gateway, relaunch blocking, network revocation	Available now

The price figures come from a newspaper report, not from the companies. Until CrowdStrike publishes terms in a filing, the split between the two deals should be treated as reported rather than confirmed.

What Fiddler adds: a record of what agents decided and why

According to CrowdStrike’s acquisition releases as summarised by Stocktwits, Fiddler will add monitoring and explainability to CrowdStrike’s Falcon platform, letting enterprises trace an AI agent’s decisions, tool usage and behaviour. “Explainability” here means being able to reconstruct why a system produced a given action, not just that it did. CrowdStrike plans to sell the technology as Falcon Observe, which is meant to let companies evaluate agents before deployment, detect unintended actions while they run, and keep audit records.

Benzinga’s coverage adds detail on what is traced: an agent’s decisions, reasoning paths and interactions with other tools, with the aim of spotting behavioural drift (an agent gradually straying from how it was meant to behave), unintended actions and failures. Before production, the technology can also evaluate agents for accuracy, task completion and compliance with company policies.

CrowdStrike frames the problem as one that existing security tools miss. “AI agents don’t have to be compromised to be dangerous,” the company said in its press release, as quoted by Benzinga. The argument is that an agent working inside approved guardrails, with valid credentials and authorised tools, can still make bad decisions. Traditional systems flag suspicious access; they are not built to explain why an authorised agent did something harmful.

CEO George Kurtz called Fiddler the “flight recorder for every agent,” a way for businesses to explain what an AI system did and why when regulators, boards or risk managers ask.

What Dreadnode adds: automated attacks before launch

Dreadnode brings automated security testing of AI models, applications and agents before deployment. CrowdStrike said Dreadnode uses more than 100 known AI attack strategies, including prompt injection and jailbreaks, to find weaknesses. Prompt injection is an attack in which malicious instructions are hidden in content an AI system reads, such as a web page or a document, so that the model follows them; a jailbreak is an input crafted to get a model to ignore its safety rules.

The findings are meant to feed into Falcon Guardian, CrowdStrike’s runtime protection for AI, as security policy recommendations. In practice that would connect pre-launch testing to what gets enforced after launch: a weakness found in testing becomes a rule applied in production. The automated testing is expected to become generally available in fiscal Q4 2027.

CrowdStrike’s website already lists an “Automated AI Security Testing” offering described as testing “every AI model, application, and agent for security weaknesses before deployment,” and an “AI Observability” offering to “evaluate agents pre-deployment then observe their behavior and enforce accountability in production”. The sources do not spell out how those existing listings relate to the products planned from the two acquisitions.

How the AI Kill Switch works, and what it cannot do yet

The AI Kill Switch sits inside Falcon Guardian and gives authorised security administrators control over AI agents that act outside their intended instructions. According to CrowdStrike’s release, it can:

• stop AI agents through gateway controls, meaning the network checkpoints agents pass through to reach tools and services;
• prevent stopped agents from relaunching;
• revoke an agent’s access to the enterprise network.

Two details matter for anyone evaluating it. First, direct termination of agents on endpoints (the laptops and servers where an agent process actually runs) is planned for a later release, not included now. Today’s version cuts an agent off from the network rather than killing the process on the machine. Second, actions triggered by administrators are recorded in audit logs, which matters for organisations that must show who shut down what and when.

The Kill Switch is available now, the company said. Benzinga describes it as a way to stop unauthorised agent activity at runtime, alongside the testing and monitoring the acquisitions are meant to bring. No independent evaluation of how quickly or reliably it stops an agent appears in the sources.

Integrations with Claude Cowork, AWS and Lovable

CrowdStrike also announced three partner integrations. The coverage describes what each one is meant to do but does not say whether they are available now or at a later date.

• Anthropic’s Claude Cowork: an integration bringing Falcon Guardian protection directly into Claude Cowork’s execution environment to detect threats and potential data exfiltration, meaning data being moved out of the company without authorisation.
• Amazon Web Services: new AWS integrations, including Falcon Guardian protection through Amazon Bedrock AgentCore Guardrails, aimed at prompt injection and sensitive-data leakage.
• Lovable: a partnership to bring visibility into applications built with AI into Falcon Shield, helping security teams find publicly accessible apps, sensitive-data risks and access permissions.

The Lovable piece addresses a different risk from the agent controls: apps produced quickly with AI app builders that may be exposed to the public internet or hold sensitive data without the security team knowing about them.

Why CrowdStrike is moving into agent security

The acquisitions extend a push CrowdStrike has been making throughout 2026. In March, it launched Agentic MDR (managed detection and response run partly by AI agents) and said AI-enabled adversaries had increased operations by 89% year over year, while average eCrime breakout time fell to 29 minutes. Those figures are CrowdStrike’s own. The company’s site now lists offerings such as “Agent Security,” described as securing AI agents “from development to adoption to runtime,” and a “Shadow AI Visibility Service” to find hidden AI tools and agents across endpoint, cloud and SaaS.

On the market side, Stocktwits notes that CRWD shares were up more than 1% in premarket trading after the announcement and have risen more than 113% so far this year. It also cites Deutsche Bank’s view that the market has yet to fully appreciate AI’s potential to expand cybersecurity, while retail sentiment on Stocktwits remained “bearish”. Benzinga frames the open question for investors as whether AI security becomes a meaningful source of customer demand and recurring revenue.

What it means for teams building AI agents

For developers and founders shipping agents into enterprises, the announcement points to three expectations that buyers’ security teams are likely to bring:

1. Pre-launch adversarial testing. Dreadnode’s approach, running a library of more than 100 known attack strategies before deployment, is the kind of check an enterprise customer may run against an agent product. Agents that read untrusted content are the most exposed to prompt injection.
2. Traceable decisions. Falcon Observe is meant to record decisions, reasoning paths and tool calls. Agents that expose clear logs of what they did and which tools they called will be easier to fit into that kind of monitoring.
3. An off switch outside the agent. The Kill Switch works at the gateway and network level, not by asking the agent to stop. Builders should assume an administrator can cut an agent’s network access mid-task and design for safe failure when that happens.

The limits are worth stating plainly. The two acquired technologies are not part of CrowdStrike’s products yet and are planned for fiscal Q4 2027. The Kill Switch does not yet terminate agents directly on endpoints. The deal values are reported, not disclosed. And the sources include no third-party tests of any of these products; every capability described here is CrowdStrike’s own claim.

What to watch next

• Deal closing and terms: whether CrowdStrike confirms the reported $400 million and $95 million figures when the transactions close, expected in fiscal Q4 2027.
• Falcon Observe and automated testing launches: both are planned for general availability in the same quarter.
• Endpoint termination for the Kill Switch: the later release that would let administrators stop agents directly on the machine.
• Integration timing: when the Claude Cowork, AWS Bedrock AgentCore Guardrails and Lovable integrations become usable, which the coverage does not say.
• Customer uptake: whether enterprises pay for agent monitoring and accountability tools, the milestone Benzinga flags as important.