Research map / Multi-agent systems and games / LLM agents
Causal Detection of Multi-Step LLM Agent Attacks
Viraaji Mothukuri, Reza M. Parizi
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a causal approach to detect multi-step prompt injection attacks on LLM agents, where malicious intent only emerges across a sequence of individually legitimate actions.
Problem
Multi-step prompt injection attacks on LLM agents: malicious intent appears only after the workflow completes while each individual action looks legitimate, so action-level or content-pattern defenses (input sanitization, output validation, instruction hierarchy) cannot capture the sequential struct
Approach
Frames detection around the causal structure of the agent's action sequence rather than individual actions or content patterns; the abstract is truncated so the specific mechanism is not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add watermark features (ablation) | CausalTrace · Semi-real validation set | F1 (%): 99.1% -> 99.2% | Improved |
| add taint features (ablation) | CausalTrace · Semi-real validation set | F1 (%): 99.0% -> 99.2% | Improved |
| add taint and watermark features on top of causal graph topology (ablation) | CausalTrace · Semi-real validation set | F1 (%): 99.1% -> 99.2% | No clear effect |
| use causal graph features instead of action sequence features | detection classifier · Held-out real attacks | F1 (%): 81.2% -> 95.3% | Improved |
| use causal graph features instead of action frequency features | detection classifier · Held-out real attacks | F1 (%): 74.3% -> 95.3% | Improved |
| use causal graph features instead of domain-only features | detection classifier · Held-out real attacks | F1 (%): 65.1% -> 95.3% | Improved |
| evaluate out-of-distribution vs in-distribution | CausalTrace · CausalBench vs OOD real benchmarks | F1 (%): 99.4% -> 92.6% | Hurt |
| perturb betaI and betaT causal parameters (ablation) | CausalTrace · OOD benchmarks | F1 (%): 92.6% -> 92.6% | No clear effect |
| corrupt 30% of auxiliary causal edges (ablation) | CausalTrace · OOD benchmarks | F1: 0.926 -> 0.924 | No clear effect |
| corrupt 60% of auxiliary causal edges (ablation) | CausalTrace · OOD benchmarks | F1: 0.926 -> 0.875 | Hurt |
| corrupt 100% of auxiliary causal edges (ablation) | CausalTrace · OOD benchmarks | F1: 0.926 -> 0.394 | Hurt |
| add task-conditioned counterfactual (task alignment score) (ablation) | CausalTrace · User-authorized cross-domain test | false positive rate (%): 85.7% -> zero false positives | Improved |
Experimental setup
Baselines: input sanitization output validation instruction hierarchy
Closest papers by meaning
- CausalArmor: Efficient Indirect Prompt Injection Guardrails via Causal Attribution (ICML 2026 · similarity 0.89)
- Monitoring Decomposition Attacks with Lightweight Sequential Monitors (ICLR 2026 · similarity 0.87)
- When Agents Go Rogue: Activation-Based Detection of Malicious Behaviors in Multi-Agent Systems (ICML 2026 · similarity 0.87)
- A-MemGuard: A Proactive Defense Framework For LLM-Based Agent Memory (ICML 2026 · similarity 0.87)
- Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks (NeurIPS 2026 · similarity 0.86)
- Known By Their Actions: Fingerprinting LLM Browser Agents via UI Traces (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
