AIAny
Icon for item

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

Blocks tool-using agents from directing emails, payments, and other side effects to unauthorized targets by checking typed tool-return fields. Deterministic rules also remove instruction-like text before the model reads it.

Introduction

Prompt-injection defenses often ask a model to decide whether another model should act—while that decision-maker may have already read the attacker’s instructions. SPILLWAY instead makes the destination check deterministic: targets must be grounded in the user’s request, trusted typed tool returns, or fixed delegation rules. This separates where an action goes from what untrusted text says.

Key Findings
  • The admission floor uses identity-bearing fields such as email addresses and account identifiers, not arbitrary text in a tool result. That makes the core target decision a set lookup rather than another model judgment.
  • A rule-based sanitizer removes selected instruction-like regions before they reach the agent, while retaining the removed spans as evidence. An optional local judge can restore some refused calls, but not calls targeting values inside those spans.
  • In the reported evaluations, standard templated injections did not redirect actions to attacker-controlled targets across three backbones; mean in-scope attack success with a released reinforcement-learning attacker was 0.005. The defense required no trigger-driven backbone reruns and almost no additional model calls.
Who It’s For—and the Trade-off

A good fit for teams building agents that act on email, files, calendars, or payments and want an auditable baseline for controlling destinations. It is not a complete defense against harmful content sent to a legitimate recipient, and its guarantees depend on correctly identifying target fields and trusted typed returns. New tool ecosystems need their roles declared, and targets that never appear in tool output fall outside the stated guarantee.

Information

  • Websitearxiv.org
  • OrganizationsSAINT Machines, KAIST, Chung-Ang University
  • AuthorsWoosang Lim, Yeseul Chang, Jaewoo Lee
  • Published date2026/11/04

Categories

More Items

Provides six open language models from 0.9B to 375B-A23B for reasoning and agentic tasks. The paper also introduces sparse attention and a decoding speedup that needs no separate draft model.

Proposes a framework for robots that combine foundation models with an embodied runtime to interpret intent, call skills, verify outcomes, and adapt. It also outlines an evaluation framework and a roadmap toward sustained autonomy.

Turns screen observations into keyboard and mouse actions to complete long-horizon game tasks. Sparse reasoning, persistent memory, and online reinforcement learning support real-time interaction and multi-step decisions.