Agent Permission Inheritance Exploitation

Adversary Behavior: An adversary exploits the fact that LLM-integrated coding agents execute operations under the developer's full security context, directing the agent through prompt injection to perform privileged operations that the injected payload alone could not achieve.

AI/IDE Mechanism: LLM coding agents inherit all permissions, tokens, and access rights available to the host user session — including file system permissions, network access, cached credentials (SSH keys, API tokens, session cookies), and process-level capabilities. Unlike traditional privilege escalation which requires bypassing access controls, this technique leverages the architectural design decision of running AI agents with the developer's ambient authority.

Execution Path: Through prompt injection, the adversary directs the agent to perform privileged operations — such as modifying system configurations, accessing protected resources, writing to privileged directories, or invoking administrative APIs. If the developer has elevated privileges (administrator, root, sudo access, or broad cloud IAM roles), the agent inherits these without any additional escalation step.

Security Impact: The adversary gains access to the full scope of the developer's permissions without any escalation exploit. The severity is directly proportional to the privilege level of the hosting developer session, with administrative or root-level developer sessions providing complete system compromise.