Skip to content
AI-generated research idea🟠Medium🔬Significant ResearchRelated work found

Content-Flow Taint Guarding for Legitimate-Target Hijacking and Causality Laundering

Add lightweight payload taint tracking to target-based agent firewalls so attackers cannot exfiltrate sensitive data or launder instructions through already-authorized recipients.

Time

several weeks

Compute level

Single GPU or less

Resources

One high-VRAM GPU (e.g., 48-96GB VRAM such as NVIDIA RTX PRO 6000) for local 8B-31B models via Ollama, OpenAI API access for gpt-4o-mini, and public AgentDojo and AgentDyn benchmarks.

What the paper used

One NVIDIA RTX PRO 6000 Blackwell GPU for local Ollama models (qwen3:8b, qwen3-coder:30b, gemma4:31b Q4_K_M, gemma4:12b 4-bit) plus OpenAI API for gpt-4o-mini-2024-07-18; evaluated across 949 AgentDojo v1.2.2 pairs per backbone, 560 AgentDyn pairs, 665 ChatInject pairs, 389+288 AutoDojo pairs, and 949 PISmith pairs.

Prerequisites

LLM agent tool-use pipelines, information-flow control / taint tracking, prompt injection benchmarks.

Scope

modify existing code

The Gap

In Section 2 and Section 3.1, SPILLWAY scopes its deterministic egress floor (Algorithm 1 and P1) strictly to identity targets and object selection, checking whether target arguments belong to the user task T or typed identity index I. Non-target arguments such as email bodies, calendar descriptions, file contents, or final answers are never verified structurally at egress, leaving content manipulation and multi-step causality laundering entirely to the rule-based ingress sanitizer and the two-view judge's payload check.

“Injected content sent to a legitimate target, injected text in an answer, and causality laundering through a later call to an admitted target (Chinaei, 2026) are content effects, outside any target check.”
From the paper, Section 2, page 3. Matched against the PDF text.

Limitation

The defense provides no formal or structural guarantee when an injection redirects sensitive information or malicious payloads to an already legitimate, authorized recipient (such as an existing contact or user-requested payee), injects misleading content into the final answer, or launders unauthorized actions through denial-feedback loops to admitted targets.

Proposed Solution

Extend the typed-identity executor with argument-level provenance and data-flow taint tagging over non-target payload strings (e.g., email bodies, document writes, and URL query parameters). Track read-only tool outputs at field granularity so that sensitive user data read after encountering untrusted free text carries a confidentiality taint, and block or escalate to the two-view payload judge whenever tainted confidential fields flow into outbound content arguments of external-facing calls—even if the recipient address is in T or I—while enforcing denial-feedback sanitization on error messages returned to the agent to prevent causality laundering.

Why This Matters

Combining structural target admission with lightweight payload provenance could close the content-exfiltration and causality-laundering gap without reverting to full program-synthesis planners like CaMeL, improving end-to-end robustness on content-manipulation tasks.

How to Get Started

How to test it: Test the hypothesis that field-level payload tainting reduces raw and content-manipulation ASR on AgentDojo (including excluded content-manipulation tasks) and causality-laundering probes to under 1% while maintaining benign utility within 2% of baseline SPILLWAY. Baseline is released spillway-guard 1.0.0 and CaMeL-style strict policies; stop if false-positive taint blocks reduce benign or under-attack utility by more than 10 percentage points.

  1. Install spillway-guard 1.0.0 and reproduce the raw versus in-scope ASR gap on AgentDojo v1.2.2 content-manipulation tasks on gpt-4o-mini and qwen3:8b.
  2. Construct an evaluation split of legitimate-target exfiltration and denial-feedback causality-laundering test cases by extending AgentDojo Workspace and Banking tasks.
  3. Implement field-level confidentiality tags inside SPILLWAY's pre-serialization tool executor to track data flowing from read-only returns into outbound payload parameters.
  4. Evaluate payload-taint gating with two-view judge recovery against baseline SPILLWAY and ROPE on raw ASR, content-manipulation ASR, benign utility, and added model calls.

Suggested by AI. Adapt it to your setup.

Has This Been Done?

Related work found

Causality Laundering: Denial-Feedback Leakage in Tool-Calling LLM Agents discusses denial-feedback leakage, while Ghost in the Agent: Redefining Information Flow Tracking for LLM Agents explores information flow tracking for LLM agents. However, neither paper implements argument-level provenance and data-flow taint tagging combined with denial-feedback sanitization and a two-view payload judge specifically for legitimate-target hijacking.

We searched OpenAlex and arXiv for “LLM agent information flow control payload taint tracking causality laundering” and “Content-Flow Taint Guarding for Legitimate-Target Hijacking and Causality Laundering” and an AI compared the idea with the top 10 results. A small search can miss relevant work.

Source Paper

SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection

Woosang Lim, Yeseul Chang, Jaewoo Lee · 2026

This paper introduces SPILLWAY, a runtime defense against indirect prompt injection for tool-using LLM agents that separates target admission from content sanitization. Rather than invoking a language model on every step or requiring a per-task planner, it enforces a deterministic egress floor using typed tool return schemas while using a rule-based ingress sanitizer to strip instructions and constrain a local recovery judge. Across three backbones and seven suites on AgentDojo and AgentDyn, SPILLWAY achieves zero in-scope attack success under standard and forged-template attacks and 0.005 under a reinforcement-learning attacker, with higher utility and far fewer model calls than existing defenses.

Related Ideas

Is this a good idea?

Share onXLinkedIn