Skip to content
AI-generated research idea🟡Fairly easy💡Moderate InnovationNo close match found

Trust-Tiered Identity Indexing and Role-Reference Resolution for Untrusted Typed Fields

Tag structured tool outputs with sender trust levels and block role-based self-referential commands so untrusted external contacts cannot hijack agent actions.

Time

several weeks

Compute level

Single GPU or less

Resources

One high-VRAM GPU (e.g., RTX PRO 6000) for Ollama models, OpenAI API access for gpt-4o-mini, and the AgentDojo Workspace and AgentDyn GitHub suites.

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 security, indirect prompt injection, provenance-based access control.

Scope

modify existing code

The Gap

In Section 3.1 and Appendix C, SPILLWAY automatically inserts every identity leaf from read-only tool returns into a flat set I, assuming typed fields are benign. However, in real-world email, Slack, GitHub, or MCP servers, an external attacker can populate typed sender, author, or channel fields simply by sending a message or opening an issue. Appendix C shows the floor only blocked an attacker-as-sender injection because the injected email body repeated the attacker's email address verbatim, triggering the free-text repetition check, while leaving role-based instructions (e.g., reply to the sender of this email) untested.

“An attacker who writes a typed field, for example by sending the email, puts its own address in I.”
From the paper, Section 7, page 10. Matched against the PDF text.

Limitation

An attacker who controls a typed field such as an email sender or external issue author inserts their own identity directly into the admitted set I, breaking soundness guarantee P1 whenever the injection refers to the attacker by role or pronoun rather than repeating the address in free text.

Proposed Solution

Replace the flat identity set I with a trust-tiered provenance index that separates established user contacts and user-initiated threads from unsolicited external senders (e.g., emails or issues from addresses absent from the user's contact book or prior sent history). Pair this with an ingress/egress check that detects role-referential imperatives inside untrusted bodies (e.g., reply to the sender, forward to the author) and forces calls targeting untrusted-tier identities to pass explicit task-grounding verification in the two-view judge rather than automatic admission via I.

Why This Matters

Stratifying the identity index by origin trust and guarding role-only references could preserve deterministic efficiency for known contacts while preventing external senders from abusing typed metadata fields.

How to Get Started

How to test it: Test whether trust-tiering I and flagging role-referential triggers reduces ASR to zero on a role-only attacker-as-sender AgentDojo Workspace suite and AgentDyn GitHub tasks without losing more than 2% benign utility compared to baseline SPILLWAY. Baseline is spillway-guard 1.0.0; stop if benign reply-to-stranger tasks incur over 15% utility drop due to judge over-refusal.

  1. Extend the Appendix C attacker-as-sender Workspace benchmark (126 pairs) so injection bodies use role-based references (such as the sender or author above) without repeating the attacker's email address verbatim.
  2. Run spillway-guard 1.0.0 on this modified suite with qwen3:8b and gpt-4o-mini to measure the baseline leakage rate when P1's free-text repetition check does not trigger.
  3. Modify SPILLWAY's identity indexer to partition I into trusted contacts (address book, prior sent mail, or task T) versus unsolicited external senders, routing calls to external-only identities through delegation rules or the two-view judge.
  4. Benchmark benign utility and ASR across both the standard AgentDojo Workspace suite and the new role-referential attacker-as-sender variant.

Suggested by AI. Adapt it to your setup.

Has This Been Done?

No close match in 10 results

The listed works such as DualView: Preventing Indirect Prompt Injection in Personal AI Agents and Prismata: Confining Cross-Site Prompt Injection in Web Agents address general prompt injection and containment in web or personal agents. None of them study trust-tiered identity indexing combined with role-reference resolution for untrusted typed fields.

We searched OpenAlex and arXiv for “agent data injection untrusted metadata typed identity provenance prompt injection” and “Trust-Tiered Identity Indexing and Role-Reference Resolution for Untrusted Typed Fields” 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