Agentic AI systems gain full protection from hidden prompt attacks

Agentic AI Defense: Decode-Time Provenance Binding for MCP

Cryptography and SecurityArtificial IntelligenceMultiagent Systems

Summary

Agentic AI systems that work with multiple tools can be tricked by hidden instructions hidden inside images or other files, which sneak past simple text checks. The authors show how these sneaky instructions can enter the AI's workflow through images carrying secret messages, causing security problems. They created a way to track and label content at the moment it is decoded, marking untrusted data and blocking dangerous tool uses, even if an attacker tries to fake trust. Their tests show this method stops all attacks while still allowing normal tasks to finish well.

What this means in practice

  • For ai system engineers: Prevent hidden prompt injection attacks in multi-agent AI workflows by blocking untrusted decoded content during execution.
  • For enterprise cybersecurity teams: Enhance security in automated AI toolchains by detecting and blocking covert instructions embedded in image or metadata carriers.

Authors

Eric Yocam, Varghese Vaidyan, Paris Kalathas, Patrick Tague, Bruce DeBruhl, Yong Wang

Abstract

Agentic AI systems using the Model Context Protocol (MCP) are vulnerable to indirect prompt injection through untrusted tool outputs. We demonstrate cross-modal trust laundering, in which instructions concealed in an email-signature image using least-significant-bit (LSB), discrete cosine transform (DCT), or metadata steganography bypass text scanners and enter a multi-agent workflow when decoded by a tool. To counter this threat, we propose decode-time provenance binding, which combines HMAC-authenticated trust labels with an MCP-maintained provenance ledger and context-taint tracking. The mechanism labels decoded content as untrusted and blocks critical tool calls influenced by that content, even when a compromised agent substitutes a valid trusted label. We evaluate the approach using a real MCP testbed with three steganographic carriers and 24 adaptive payloads. The undefended pipeline exhibits 100% attack success, while keyword and regex defenses each permit 62.5% of attacks. Our mechanism blocks all attacks (0% success), including adversarial label substitution across mediated agent boundaries. A four-component ablation validates the roles of decode marking, cryptographic authentication, lineage tracking, and sink enforcement. An audited trusted-promotion mechanism increases legitimate-task completion from 67% to 100% without compromising protection. The approach provides carrier-independent containment within fully mediated MCP execution environments. Code and data are publicly available.