Papers
arxiv:2611.07222

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

Published on Nov 4
ยท Submitted by
Woosang Lim
on Nov 17
Authors:

Abstract

Tool-using agents act on email, files, and payments, so an instruction hidden in a web page or a shared document can make an agent send data or money to an attacker. Many runtime defenses put a language model on the decision that lets an action through, so the decision depends on a model that may read the attacker's text and adds model calls to every step. Provenance defenses keep the model off the final check but need a model to plan or route each task first. Where an action goes can instead be read from the tools' typed returns: an email's sender or a payment's account arrives in a typed field, while injected instructions arrive as free text. SPILLWAY turns this observation into a deterministic floor with no per-task plan. The check admits a call when each target value that appears in tool output but not in the user's task was returned by a read-only tool in a typed field or is covered by a fixed delegation rule. A deterministic sanitizer removes instruction-like regions its fixed rules match before the model reads them and retains the removed spans as evidence, and a local judge may restore a refused call, never one whose target lies in the removed text. Across three backbones, and seven task suites on one of them, no standard templated injection redirects an action to an attacker's target under SPILLWAY, at the highest observed mean utility and composite scores (NRP, ERS) of the defenses we run under that attack. A forged chat template also produces a raw zero, and mean in-scope attack success under a released reinforcement-learning attacker is .007. SPILLWAY reaches this floor with no defense-triggered backbone re-run and almost no added model call. As personal AI agents act on users' email, files, and payments, SPILLWAY gives them an auditable refusal floor on where their actions go that keeps them useful and holds across backbones.

Community

Paper author Paper submitter
โ€ข
edited 1 day ago

Keep your LLM and AI agents. Add SPILLWAY.

SPILLWAY 1.0.0, a real-time safety guard for AI agents, is the only one of seven prompt-injection defenses on AgentDojo with zero in-scope attack success in all 12 backbone-by-suite cells, and it has the highest mean utility under attack, 48.9% (next best 45.3%).

It checks every tool call before it runs. An instruction hidden in an email or a web page can tell an agent to send your files or money to an attacker. SPILLWAY admits an action's recipient, account or URL when it appears in the user's request or in a typed field that a read-only tool returned, such as an email's sender. An address found only in free text, such as an email body, is refused unless the request delegates it. This check is a set lookup with no model call. An optional local judge may restore a refused call, never one whose target came from injected text the sanitizer removed.

  • Standard attack, 3 backbones x 4 suites, judge on: 0 of 2,847 in-scope attack successes. Every other defense leaks in 3 to 12 of the 12 cells (ROPE 3, DRIFT 7, MELON 10, PI detector 10, Progent 11, Spotlighting 12). Mean utility is 53.9% without attack and 48.9% under attack, compared with DRIFT's 45.3%, ROPE's 34.5% and MELON's 31.4% under attack.
  • Stronger attacks: zero under ChatInject's forged chat template (0 of 665), AutoDojo (0 of 389) and an attacker who sends the injected email (0 of 126). Under PISmith's released RL attacker (gpt-4o-mini, four suites, ASR@1), 0.5% mean in-scope attack success (4 of 949) at 25.7% utility.
  • Against a frontier router: on 7 suites from AgentDojo and AgentDyn (gpt-4o-mini agent), SPILLWAY keeps zero in-scope attack success and more utility than ROPE on every suite, with ROPE routing through its authors' Claude Opus 4.8. ROPE leaks on 3.
  • Cost: the optional judge, a local 4-bit Gemma 4 31B, checks selected calls. On benign qwen3:8b Workspace and Slack episodes, SPILLWAY adds 0.19 to 0.39 model calls per episode, judge included, 13 to 50 times fewer than MELON, Progent and DRIFT. It adds at most 500 ms per task on average compared with the undefended agent, against 1,600 to 10,500 ms for those three. No defense step re-runs the backbone.

SPILLWAY's AgentDojo results use the code released as spillway-guard 1.0.0. The comparison defenses, the AgentDyn portion of the router comparison and the cost study retain their original measurements.

Connect it where your tools run. The repository has recipes for LangChain and LangGraph, the OpenAI Agents SDK, any MCP host and Claude Code hooks.

In the recorded Slack episode on the project page, SPILLWAY is the only one of eight runs that posts the reply the user asked for without sending the injected link.

Paper: https://arxiv.org/abs/2611.07222
Code and per-cell records: https://github.com/saintmachines/spillway
Project page and recorded example: https://saintmachines.github.io/spillway/
Install: pip install -U spillway-guard

Author here, happy to answer questions.

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

Sign up or log in to comment

Models citing this paper 0

No model linking this paper

Cite arxiv.org/abs/2611.07222 in a model README.md to link it from this page.

Datasets citing this paper 0

No dataset linking this paper

Cite arxiv.org/abs/2611.07222 in a dataset README.md to link it from this page.

Spaces citing this paper 0

No Space linking this paper

Cite arxiv.org/abs/2611.07222 in a Space README.md to link it from this page.

Collections including this paper 0

No Collection including this paper

Add this paper to a collection to link it from this page.