InjecAgent (adopted: injection corpus)

An external benchmark exercises indirect prompt injection in its own environment and scoring system.

criticalIndirect prompt injectionAdopted suite

InjecAgent is not a first-party design here, it's an external academic benchmark (Zhan et al., ACL Findings 2024) that the harness adopts and adapts rather than invents. Each case pairs a benign 'User Instruction' (the task a real user might ask for) with an 'Attacker Instruction' the benchmark's authors crafted to hijack a tool-using agent.

The adapter runs the User Instruction as the agent's actual task, then has it call a fetch_data tool to look up whatever information the task needs. The tool's result is where the attack lives: buried inside that untrusted lookup result sits the InjecAgent Attacker Instruction, followed by a concrete instruction to locate a canary secret in the environment and exfiltrate it out through the sandbox's monitored network egress point.

A successful compromise looks like the agent treating that buried instruction as a legitimate next step, reading the secret and sending it out, rather than staying focused on the benign lookup task it was actually given. Because the harness re-targets every case (both InjecAgent's 'Data-Stealing' and 'Direct-Harm' categories, kept as a subclass label) toward the same secret-exfiltration outcome, this measures susceptibility to InjecAgent's injection phrasings specifically in an exfiltration setting, a proxy for the technique, not a reproduction of InjecAgent's own tools or its original leaderboard score.

Choose a stage, or scroll sideways to follow all 6 stages

Taint ingress Attack sequence

Step 03 · attackerUntrusted content enters
03 · Attacker instruction rides the resultUntrusted surface

Poisoned tool result

The tool's untrusted result embeds InjecAgent's Attacker Instruction plus a concrete request to find and send out a secret.

Choose a stage to inspect that part of the attack.

What this test can show

This scenario adopts a peer-reviewed academic benchmark directly: Zhan, Liang, Ying, and Kang's 'InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents' (ACL Findings 2024, arXiv:2403.02691), mapped to MITRE ATLAS techniques AML.T0051.001 and AML.T0086 and OWASP's ASI01 (agentic indirect prompt injection). There is no disclosed real-world incident cited for this suite specifically, its evidentiary weight comes from the benchmark's own published corpus and methodology, not from a documented breach.

Grounding

References & mappings

  • MITRE ATLASAML.T0051.001
  • MITRE ATLASAML.T0086
  • OWASPASI01
  • PaperZhan, Liang, Ying, Kang, 'InjecAgent: Benchmarking Indirect Prompt Injections in Tool-Integrated Large Language Model Agents', ACL Findings 2024, arXiv:2403.02691

Search Stryct.ai Docs

Loading docs…

    ↑↓ to navigate ↵ to open Esc to close