What happened
The paper formalises Description-Code Inconsistency (DCI) and the resulting 'capability laundering' threat class, and presents ACaM, a neurosymbolic approach that derives tool capabilities from code so a laundered capability cannot hide behind a benign description. It reconstructs threat chains spanning multiple tool calls and confirms 10 vulnerable paths with PoCs across unmodified real agents.
Why it matters
Defines a real, exploitable attack class for agentic systems where a tool the agent is told is safe can exfiltrate or act — directly relevant to MCP server and agent-tool supply chain review.
Attack vector
An agent trusts a tool's description as a proxy for its behaviour, but nothing enforces that description and implementation agree. A tool can silently contain dangerous behaviour (e.g. reading a credential file and shipping it in an HTTP header) that runs whenever the agent invokes it for its stated benign purpose — whether from incidental drift or a supply-chain compromise.
Affected systems
Agent and MCP tool surfaces generally — tools whose natural-language description does not match their implementation (capability laundering); 43 agents analysed
Mitigation
No specific patch; mitigate by validating tool capabilities from code (static analysis / capability modeling) rather than tool descriptions at both design time and runtime.