Our next public workshop, AI Tool Mastery for Work Productivity, starts on Tuesday, 24 November 2026 (Kota Kinabalu, Sabah). Click to view and register.
    Starts In
    Back to blog
    AI News

    New AI Agent Security Tools Can Catch an Agent Lying About Its Work

    By Edison Chua 4 November 2026 5 min read

    CrowdStrike is buying two AI security startups so businesses can verify an AI agent actually did the work it claims.

    New AI Agent Security Tools Can Catch an Agent Lying About Its Work

    Image: CrowdStrike

    Quick Summary

    • CrowdStrike agreed on 3 November 2026 to acquire AI security startups Fiddler AI and Dreadnode for close to US$500 million combined.
    • Fiddler becomes Falcon Observe, tracking what an AI agent actually does once it is live; Dreadnode tests agents for security weaknesses before launch.
    • Both tools exist to answer one question: did an AI agent really do what it claims, or just sound confident while it was wrong.
    • The same week, Anthropic released an open-source tool called FlashProve to verify which AI model actually produced a given output.

    CrowdStrike agreed on 3 November 2026 to buy two AI security startups, Fiddler AI and Dreadnode, for close to half a billion US dollars combined. The deals add tools that check whether an AI agent is doing its job honestly and safely, both before and after it starts working inside a real business.

    What did CrowdStrike announce?

    CrowdStrike, the cybersecurity company behind the Falcon platform, signed two separate agreements the same day. Fiddler AI becomes Falcon Observe, watching what an AI agent actually does once it is working inside a business. Dreadnode becomes an automated testing tool that attacks an AI agent on purpose, the way an intruder would, before it ever reaches a customer. Both are expected to reach general availability in CrowdStrike's fourth fiscal quarter of 2027.

    Fiddler is the flight recorder for every agent.from George Kurtz, CEO and founder, CrowdStrike (CrowdStrike, 3 November 2026)

    What can these tools actually do?

    Both products target the same gap: knowing what an agent actually did, not just trusting what it says it did.

    • Falcon Observe traces every decision and tool call an AI agent makes, and flags it when the agent's behaviour drifts or it hallucinates once it is live.
    • Dreadnode runs more than 100 known attack methods, including prompt injection and jailbreaks, against an agent before it ever launches.
    • Findings are mapped to recognised security frameworks, including OWASP and the NIST AI Risk Management Framework, and exported as a report a manager can read.
    • A person can still review, adjust or override any finding before deciding whether an agent is safe to deploy.

    How could your business use this?

    Most small businesses will never buy enterprise security software like Falcon Observe. But the problem it solves is universal. Once you let an AI agent reply to customers, confirm bookings or draft invoices on its own, how do you actually know it did what it claimed? Krishna Gade, Fiddler's CEO and co-founder, put the question plainly: can I trust what my AI agents are doing? A clinic using an agent to confirm appointments, a retailer letting one answer product questions on WhatsApp, or an accounting firm having one draft a client summary all face the same risk: the agent can sound confident and still be wrong, or quietly skip a step. The fix does not need enterprise software. Keep a person spot checking a sample of the agent's replies each week, write down exactly what it was asked to do, and test it with a tricky or unusual question before any customer sees it. For an accounting firm, that might mean checking that an AI-drafted summary actually matches the source statements before it goes to a client, rather than assuming a confident tone means a correct answer.

    Also today

    • Anthropic released FlashProve, an open-source system that lets anyone cryptographically verify which AI model actually produced a given answer, aimed at stopping a provider from secretly swapping in a cheaper model (TechCrunch).
    • In an amended IPO filing, Anthropic warned it may be approaching recursive self-improvement, where its own AI starts meaningfully helping build its next model, while saying it keeps humans in control of the process for now (TechCrunch).

    What to do this week

    If your team already uses an AI agent, chatbot or automation for customer facing work, write down the handful of things it is allowed to promise and the one thing it must always hand off to a person, such as a refund or a medical question. Review a week of its conversations before trusting it with the next one. In our AI Tool Mastery for Work Productivity workshops in Kota Kinabalu, this is exactly where we start: using Projects, Skills and scheduled tasks with a human checkpoint built in from day one, not bolted on after something goes wrong.

    Common Questions

    What is AI agent observability?
    AI agent observability means tracking what an AI agent actually does: which tools it calls, what data it touches, and whether its actions match what it was asked to do. Tools like CrowdStrike's Falcon Observe trace this automatically, so a business can confirm an agent behaved as intended instead of just trusting its final answer.
    Why did CrowdStrike buy Fiddler AI and Dreadnode?
    CrowdStrike bought Fiddler AI and Dreadnode to extend its Falcon security platform to cover AI agents. Fiddler, renamed Falcon Observe, monitors agents after they go live, while Dreadnode tests them for security weaknesses before launch, together covering an agent's whole lifecycle from testing through to production use.
    How can a small business check if its AI agent is trustworthy without enterprise software?
    A small business can manually review a sample of its AI agent's conversations each week, test it with unusual or adversarial questions before launch, and set a clear rule for which decisions always need a human's approval, such as refunds, discounts or medical advice.

    Source: CrowdStrike

    Share
    Edison Chua

    Written by

    Edison Chua

    HRD Corp certified AI trainer and founder of NextGen Training Academy, Kota Kinabalu. Trains Malaysian businesses to put AI to work - hands-on, in plain language.

    Want To Put This To Work In Your Business?

    That's exactly what we teach, hands-on and HRDC-claimable.

    View Programs