CrowdStrike launches tools to detect risky AI agents

CrowdStrike has launched two security products aimed at AI agents and AI-driven attacks, expanding its Falcon platform into what it describes as agentic security operations.

The first, Agent Insider Threat Detection, is designed to identify unusual behaviour by AI agents inside an organisation. The second, Agentic Adversary Emulation, is intended to simulate how attackers could use autonomous agents to chain together weaknesses and reach critical systems.

The launches reflect a broader shift in cyber security, with AI agents treated not only as tools for defenders but also as potential insider risks and attacker instruments. CrowdStrike is positioning both products around the idea that automated threats can move too quickly for manual investigation and response.

Agent Insider Threat Detection is being added to Falcon Next-Gen SIEM. The system learns the normal behaviour of each AI agent and the person operating it, then flags deviations even when no formal policy has been broken.

Those alerts feed into coordinated investigations across endpoint, identity, software-as-a-service, cloud and network environments. These investigations are now generally available and are designed to produce a verdict within seconds, with a Falcon Guardian AI Kill Switch able to terminate an agent immediately.

The focus addresses a specific problem raised by wider adoption of autonomous software in large organisations. As companies give agents direct access to systems and sensitive data, security teams face the prospect of software acting beyond its remit without triggering conventional rule-based alarms.

"AI agents can be both an insider threat and an adversary," said Michael Sentonas, President, CrowdStrike.

"One acts without authorisation from inside your walls. The other is weaponised against you from outside. Stopping the insider means knowing the moment an agent turns, investigating across every domain, and shutting it down at machine speed. That's the next evolution of the agentic SOC," Sentonas said.

Broader controls

CrowdStrike also outlined related additions planned for the same area. These include detections for attacks on AI systems such as large language model API abuse, stolen API keys and compromises involving coding assistants, MCP servers and agent frameworks.

It is also extending agent-led triage to other parts of the Falcon portfolio, including Falcon Guardian, cloud security, SaaS security and data security. Planned additions include approval controls for whole agents, applications and workflows, alongside tools to monitor and limit AI credit consumption by security agents and workflows.

Another planned addition is a policy-drift safeguard designed to screen interactions involving Charlotte AI and AgentWorks for unauthorised tool use and access outside an agent's intended scope. CrowdStrike is also expanding federated search connectors so teams can query agent data across third-party environments without re-ingesting it.

Attack simulation

The second announcement centres on offensive testing. Agentic Adversary Emulation, powered by CrowdStrike's SafeMind models, is designed to behave like an automated attacker inside a customer-approved testing environment.

The product can chain vulnerabilities, misconfigurations and identity weaknesses into attack paths leading to critical assets. It then recommends remediation steps and re-runs the attack sequence to check whether each path has been closed.

The system draws on Falcon sensor telemetry, threat intelligence and incident response knowledge, combined with each customer's environmental context. Testing is informed by intelligence on more than 290 tracked adversaries, as well as newly disclosed vulnerabilities and emerging AI-driven threats.

That pitch places the product in a market where security providers are trying to move beyond periodic penetration tests towards continuous assessment. CrowdStrike argues that point-in-time testing cannot keep pace with attackers using autonomous systems to probe, adapt and exploit weaknesses.

"The agentic adversary uses AI to operate at a speed and scale no human can match. These are preventable gaps that go unprevented," said Adam Meyers, Head of Counter Adversary Operations, CrowdStrike.

"Now defenders can find and close them before the adversary does. That is offense in the service of defence," Meyers said.

Availability

Some of the new functions are available now, including coordinated multi-agent investigations and the broader automation workspace for agents, workflows and custom applications. Agent Insider Threat Detection and several related features are listed for release in the fourth quarter, while the policy-drift guardrails are slated for the first quarter of the next financial year.

Agentic Adversary Emulation is being offered in early access to select customers through Project QuiltWorks partners. The service is intended to let customers attack, remediate and validate on the same platform.

The announcements show how security vendors are adapting to a world in which AI agents can be operators, targets and attack tools at the same time. CrowdStrike's argument is that security teams will need automated monitoring and automated testing to handle software that can move, decide and act without waiting for a human analyst.