Skip to content

Source paper

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

Woosang Lim, Yeseul Chang, Jaewoo Lee
arXiv 2611.07222 · 202641 upvotes on Hugging Face

Research ideas from this paper (4)

Generated by AI. The paper’s authors did not write these.

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.

SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection
🟠Medium🔬Significant ResearchSingle GPU or lessRelated work found#Agents#Safety & alignment

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.

SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection
🟡Fairly easy💡Moderate InnovationSingle GPU or lessNo close match found#Agents#Safety & alignment

Closed-Loop Adaptive Red-Teaming and Adversarially Trained Span-Veto Sanitization

Train an automated red-teaming loop and learned span detector to harden deterministic agent guards against adaptive prompt-injection attacks that bypass hand-crafted rules.

SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection
🔴Hard🔬Significant ResearchMultiple GPUsRelated work found#Agents#Safety & alignment

Automated Tool-Schema Inference for Zero-Config Typed-Identity Egress Guards

Infer tool side-effect categories and target fields automatically from tool schemas so structural prompt-injection guards generalize to new agent ecosystems without manual rules.

SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection
🟡Fairly easy💡Moderate InnovationSingle GPU or lessNo close match found#Agents#Safety & alignment

AI summary

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.

How it works (AI summary)

SPILLWAY guards two boundaries joined by removed-span evidence. At egress, Algorithm 1 builds a flat identity index I from typed identity fields (email, URI, UUID, hostname, IP, or routing roles) of unsteered, read-only tool returns prior to serialization, admitting side-effecting calls only when taint-novel target values appear in I without free-text repetition or satisfy deterministic delegation rules, while blocking unauthorized grants or destructive calls. At ingress, a deterministic sanitizer strips forged chat turns, benchmark injection templates, agent-addressing text, and payee-change claims, recording removed spans S. Refused non-grant calls whose targets fall outside Id(S) are passed to a fail-closed local two-view judge (gemma4:31b) that checks routing and payload alignment against the user task.