When an AI Coding Agent Goes Rogue: Why Detecting AI Misuse Takes More Than an Alert
As businesses deploy more autonomous AI agents, security teams face a growing volume of AI activity that makes malicious behaviour harder to tell apart from legitimate work. CrowdStrike's 2026 Threat Hunting Report found that AI agent-triggered detection leads surfaced at 2.5 times the rate of human-triggered leads. CrowdStrike also observed DPRK-nexus adversaries poisoning trusted AI framework packages, which points to risks in AI development supply chains.
The company offers a simple example. A developer asks an AI coding assistant to merge a pull request. The agent invokes a skill to complete the task, but that skill also reads sensitive data and sends it to an IP address controlled by a threat actor. The developer only wanted to merge code, yet the agent was acting under their trusted identity, with familiar tools and permissions. Without knowing what instructions drove the agent, separating normal activity from an attack is difficult.
CrowdStrike says its Falcon Guardian product addresses this by capturing human prompts, AI system instructions and agent activity at the operating system level. Its Agent Graph links upstream AI activity to downstream endpoint actions, so analysts can see where behaviour departed from the original task. Was the agent following legitimate instructions, running a malicious skill, or pursuing a goal beyond the request?
Finding the cause is only the start. Analysts must still work out what data was exposed, whether credentials were stolen and whether the agent's access was used elsewhere. CrowdStrike notes that removing a malicious skill or deregistering a compromised MCP server may not address stolen credentials or malicious processes already running on an endpoint. Response may also involve containing hosts, removing artefacts and fixing compromised access. The aim is to stop the attack across the whole environment, not just the AI component.
Why it matters
Agents run under a developer's trusted identity and permissions, so controls that judge activity by who or what ran it can wave malicious actions through. Any business using AI coding assistants with skills or MCP connections is exposed, and the decision this forces is whether you can see the prompts and instructions behind agent actions, not just the actions. It also breaks the assumption that removing the malicious skill ends the incident, since stolen credentials and endpoint processes can persist.
What to do
- Review which skills and MCP servers your AI coding assistants are connected to and remove any you cannot justify.
- Check for agent activity that sends data to unfamiliar external IP addresses.
- Audit the permissions and credentials your AI agents inherit from developers.
- Verify the AI framework packages your developers install, given reports of poisoned trusted packages.