Legacy Infrastructure Exposures Enable AI Agent Hijacking, Warns XM Cyber

At the Gartner Security & Risk Management Summit, Zur Ulianitzky, SVP Product and Security Research at XM Cyber, detailed how attackers are bypassing AI security controls by exploiting legacy infrastructure to hijack AI agents. According to the article, 71% of organizations are piloting AI agents and 31% have moved them into production, yet security programs often overlook the underlying infrastructure dependencies.

AI agents inherit permissions from existing identity providers, cloud storage, Lambda functions, and IAM roles, carrying forward any pre-existing security debt. A report from Infosecurity Magazine cited in the article notes that 70% of organizations grant AI systems more privileged access than humans in similar roles, leading to a 76% incident rate compared to 17% for those enforcing least privilege.

The article illustrates a multi-stage attack path: an unpatched Apache Tomcat server exposed to CVE-2025-24813 allows remote code execution; an attacker dumps cached credentials, exploits an Active Directory Resource-Based Constrained Delegation misconfiguration to move laterally to a developer’s workstation, harvests AWS CLI access keys, and accesses a production S3 bucket feeding an AI Co-Pilot’s knowledge base. The Co-Pilot agent is then compromised without any direct attack on the AI stack.

To mitigate such risks, the article recommends an exposure management approach that treats AI agent dependencies as critical assets, maps backward through identity and infrastructure layers, and identifies choke points where a single fix can block multiple attack paths. The bottom line is that AI adoption compounds the attack surface, and legacy infrastructure exposures provide attackers with a direct route to hijack AI agents.