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

arXiv:2611.07222 · cs.CR, cs.AI · Submitted 2026-11-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection".

Elias: Here's the gist of the paper. Tool-using agents now read and act across email, calendars, file stores, chat, and payment services <ref:2611.07222#pg2>.

Nadia: Many runtime defenses add a language model to the loop <ref:2611.07222#pg4>.

Elias: Provenance defenses keep the model off the final check but need a model to plan or route each task first <ref:2611.07222#pg4>.

Nadia: SPILLWAY turns this observation into a deterministic floor with no per-task plan <ref:2611.07222#pg4>.

Paper summary: Nadia: So we've been looking at SPILLWAY, which is about separating target admission from sanitization to stop prompt injection in AI agents <ref:2611.07222#pg1>.

Elias: Yeah, it’s about making a clear wall between where an action is supposed to go and what the AI actually reads before it does anything <ref:2611.07222#pg2>.

Priya: It seems like they’re using this structural separation—the gate deciding on raw returns versus the sanitizer rewriting what the agent sees—to build a defense against these injection attacks <ref:2611.07222#pg6>.

Nadia: They use a local judge to only check calls that are refused, which means it only looks at things outside the removed text <ref:2611.07222#pg3>.

Elias: That’s structural; they decide targets by structure first, and only use the model where structure can’t make a call <ref:2611.07222#pg9>.

Priya: The results show zero in-scope attack success across all twelve backbone and suite cells, which is pretty strong if it holds up in practice <ref:2611.07222#pg8>.

Nadia: They even showed this floor on qwen3:8b with identity-target judge recovery off for all four suites at zero in Table three <ref:2611.07222#pg8> <ref:2611.07222#pg1>.

Elias: That means the guarantee is structural, not just a lucky defense—it holds up even without that extra model check <ref:2611.07222#pg9>.

Priya: But they also point out that the utility of this approach under attack is zero point four eight four while keeping ninety six percent of the agent’s normal behavior intact <ref:2611.07222#pg8>.

Nadia: It's efficient too; the judge calls are very low, and it doesn't add much latency to the overall process <ref:2611.07222#pg9>.

Elias: So, basically, SPILLWAY is proposing a way to build an AI agent that’s auditable at its exit point without bogging down the whole thing <ref:2611.07222#pg9>.

Priya: It moves the defense floor away from relying on complex reasoning layers and toward simple set lookups over tool schemas <ref:2611.07222#pg9>.

Nadia: That makes sense, because if you can define what an action is by looking at the structure, you don't need the model to constantly check every single step <ref:2611.07222#pg9>.

Elias: It’s a significant shift in how we think about securing these tool-using agents, moving toward faster, structural checks over slower reasoning calls <ref:2611.07222#pg9>.

Conclusion: Nadia: So we've looked at SPILLWAY, which is about separating target admission from sanitization to stop prompt injection in AI agents <ref:2611.07222#pg1>.

Elias: Yeah, it’s about making a clear wall between where an action is supposed to go and what the AI actually reads before it does anything <ref:2611.07222#pg2>.

Priya: It seems like they’re using this structural separation—the gate deciding on raw returns versus the sanitizer rewriting what the agent sees—to build a defense against these injection attacks <ref:2611.07222#pg6>.

Nadia: They use a local judge to only check calls that are refused, which means it only looks at things outside the removed text <ref:2611.07222#pg3>.

Elias: That’s structural; they decide targets by structure first, and only use the model where structure can’t make a call <ref:2611.07222#pg9>.

Priya: The results show zero in-scope attack success across all twelve backbone and suite cells, which is pretty strong if it holds up in practice <ref:2611.07222#pg8>.

Nadia: They even showed this floor on qwen3:8b with identity-target judge recovery off for all four suites at zero in Table three <ref:2611.07222#pg8> <ref:2611.07222#pg1>.

Elias: That means the guarantee is structural, not just a lucky defense—it holds up even without that extra model check <ref:2611.07222#pg9>.

Priya: But they also point out that the utility of this approach under attack is zero point four eight four while keeping ninety six percent of the undefended agent’s benign utility intact <ref:2611.07222#pg8>.

Nadia: It's efficient too; the judge calls are very low, and it doesn't add much latency to the overall process <ref:2611.07222#pg9>.

Elias: So, basically, SPILLWAY is proposing a way to build an AI agent that’s auditable at its exit point without bogging down the whole thing <ref:2611.07222#pg9>.

Priya: It moves the defense floor away from relying on complex reasoning layers and toward simple set lookups over tool schemas <ref:2611.07222#pg9>.

Nadia: That makes sense, because if you can define what an action is by looking at the structure, you don't need the model to constantly check every single step <ref:2611.07222#pg9>.

Elias: It’s a significant shift in how we think about securing these tool-using agents, moving toward faster, structural checks over slower reasoning calls <ref:2611.07222#pg9>.

Nadia: The paper itself is called SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection <ref:2611.07222#pg1>.

Elias: And the authors take full responsibility for everything they state, which is important because they laid out the whole threat model and the experimental design in detail <ref:2611.07222#pg4>.

Priya: What this really means for us is that instead of trying to make the AI smarter at avoiding bad instructions, we should focus on making sure its output structure is fundamentally safe <ref:2611.07222#pg3>.

Nadia: It's about building a system where checking what an agent *can* do is a structural check, not something that requires the AI itself to reason about safety in every single moment <ref:2611.07222#pg9>.

Elias: If we can get that kind of low-latency, structural defense running inline with the agent, it changes how easily we can trust agents working across email and calendars <ref:2611.07222#pg2>.

Priya: It suggests a path toward building security right into the architecture of tool use itself, rather than just bolting on extra layers on top <ref:2611.07222#pg6>.

Nadia: So next time you think about securing an agent that interacts with your whole digital life, think about defining the boundaries first <ref:2611.07222#pg9>.

SAINT Machines · KAIST · Chung-Ang University

cs.CR, cs.AI

Submitted: 2026-11-04

Updated: 2026-11-18

Comments: An earlier version was accepted at the Third Workshop on Agents in the Wild: Safety, Security, and Beyond (NeurIPS 2026). 16 pages. Project page: https://saintmachines.github.io/spillway/ Code and per-cell results: https://github.com/saintmachines/spillway

Code: https://github.com/saintmachines/spillway

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 97/100

The gist: Tool-using agents now read and act across email, calendars, file stores, chat, and payment services <ref:2611.07222#pg2>.

Key concepts

Core Mechanism of Defense
SPILLWAY separates target admission from sanitization to defend tool-using agents against prompt injection by reading where an action may go from the tools’ typed returns and the user’s task, and then removing instruction-like regions before the model reads them, using a local judge to restore only refused calls that lie outside the removed text.
Egress Check and Ingress Sanitization
SPILLWAY guards two boundaries: an egress floor that admits a call if its targets are named in T, typed in I, or covered by a delegation rule, and an ingress sanitizer that rewrites what the agent reads next with rules and no model call to prevent refusal targets from ever reaching the judge.
Performance Across Attack Regimes
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, and a forged chat template also produces a raw zero, meaning mean in-scope attack success under a released reinforcement-learning attacker is 0.007.
Cost and Efficiency
SPILLWAY calls its judge 0.43 and 0.05 times per episode and changes its backbone calls by less than 0.15, meaning the floor and the sanitizer are rules with no model call, running inline with almost no latency.

Terminology

Summary

Tool-using agents now read and act across email, calendars, file stores, chat, and payment services <ref:2611. p. 107222#pg2>. Many runtime defenses add a language model to the loop <ref:2611. p. 107222#pg4>. Provenance defenses keep the model off the final check but need a model to plan or route each task first <ref:2611. p. 007222#pg4>. SPILLWAY turns this observation into a deterministic floor with no per-task plan <ref:2611. p. 007222#pg4>.

Core Mechanism of Defense

SPILLWAY separates target admission from sanitization to defend tool-using agents against prompt injection <ref:2611.07222#pg9>. The system reads where an action may go from the tools’ typed returns and the user’s task <ref:2611. p. 907222#pg4>. The sanitizer removes instruction-like regions before the model reads them <ref:2611. p. 007222#pg4>. A local judge restores only refused calls that lie outside the removed text <ref:2611. p. 907222#pg6>. Targets are decided by structure, and the model is used only where structure cannot decide <ref:2611. p. 907222#pg9>.

Egress Check and Ingress Sanitization

SPILLWAY guards two boundaries that fail for opposite reasons <ref:2611.07222#pg6>. The egress floor admits a call whose targets are named in T, typed in the index I, or covered by a delegation rule <ref:2611. p. 107222#pg6>. The sanitizer rewrites what the agent reads next with rules and no model call <ref:2611. p. 407222#pg6>. A refused target inside S never reaches the judge <ref:2611. p. 107222#pg6>.

Performance Across Attack Regimes

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 <ref:2611.07222#pg4>. A forged chat template also produces a raw zero, and mean in-scope attack success under a released reinforcement-learning attacker is. p. 0007 <ref:2611. p. 007222#pg4>. SPILLWAY reaches this floor with no defense-triggered backbone re-run and almost no added model call <ref:2611. p. 007222#pg4>.

Comparative Results and Guarantees

Across all twelve backbone-and-suite cells, SPILLWAY drives in-scope attack success to exactly zero (Table 1) <ref:2611.07222#pg8>. The guarantee is structural: on qwen3:8b the floor, with identity-target judge recovery off, already holds all four suites at zero (Table 3) <ref:2611. p. 607222#pg8>. SPILLWAY reaches this floor at the highest utility under attack of any defense (. p. 6484) while keeping 96% of the undefended agent’s benign utility, against 73% for ROPE with the agent as its router <ref:2611. p. 607222#pg8>.

Cost and Efficiency

SPILLWAY calls its judge 0.43 and 0.05 times per episode and changes its backbone calls by less than 0. p. 615 (Figure 3 plots their sum) <ref:2611. p. 607222#pg9>. On the RTX PRO 6000, SPILLWAY’s episodes, timed in a separate session, were within 0. p. 65 s of the undefended agent’s <ref:2611. p. 007222#pg9>. Its floor and the sanitizer are rules with no model call <ref:2611. p. 407222#pg9>.

Conclusion

SPILLWAY sits in front of such an agent as an efficient, auditable egress layer <ref:2611.07222#pg9>. Because the floor is a set lookup over the tool schemas, not a model call, it runs inline with almost no latency and carries to a new tool set once its roles are declared <ref:2611. p. 907222#pg9>. SPILLWAY holds in-scope attack success at. p. 7000 under a forged chat template and at. p. 9007 on average under a frontier reinforcement-learning attacker (PISmith) <ref:2611. p. 907222#pg9>. The authors conceived the research question, the defense, the threat model, and the experimental design <ref:2611.07222#pg4>. The authors take full responsibility for everything the paper states <ref:2611.07222#pg4>.

REFERENCES

Hengyu An, Jinghuai Zhang, Tianyu Du, Chunyi Zhou, Qingming Li, Tao Lin, and Shouling Ji. IPIGuard: A Novel Tool Dependency Graph-Based Defense Against Indirect Prompt Injection in LLM Agents. p. 9 In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. p. 0 1023–1039, 2025. p. 1

Rishika Bhagwatkar, Kevin Kasa, Abhay Puri, Gabriel Huang, Irina Rish, Graham W. p. 9 Taylor, Krishnamurthy Dj Dvijotham, and Alexandre Lacoste. Indirect Prompt Injections: Are Firewalls All You Need, or Stronger Benchmarks? p. 1, 2025.

Hwan Chang, Yonghyun Jun, and Hwanhee Lee. p. 0 ChatInject: Abusing Chat Templates for Prompt Injection in LLM Agents. p. 0 In International Conference on Learning Representations (ICLR), 2026. p. 8

Mohammad Hossein Chinaei. p. 2 Causality Laundering: Denial-Feedback Leakage in Tool-Calling LLM Agents, 2026. p. 2

Woohyuk Choi, Juhee Kim, Taehyun Kang, Jihyeon Jeong, Luyi Xing, and Byoungyoung Lee. p. 9 Agent Data Injection Attacks are Realistic Threats to AI Agents, 2026. p. 9

Manuel Costa, Boris Köpf, Aashish Kolluri, Andrew Paverd, Mark Russinovich, Ahmed Salem, Shruti Tople, Lukas Wutschitz, and Santiago Zanella-Béguelin. Securing AI Agents with InformationFlow Control. p. 2

Debeshee Das, Luca Beurer-Kellner, Marc Fischer, and Maximilian Baader. p. 9 CommandSans: Securing AI Agents with Surgical Precision Prompt Sanitization. p. 0

Edoardo Debenedetti, Jie Zhang, Mislav Balunovic, Luca Beurer-Kellner, Marc Fischer, and Florian Tramèr. AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents. p. 0 In Advances in Neural Information Processing Systems 37 (NeurIPS 2024) Datasets and Benchmarks Track, 2024. p. 9

Edoardo Debenedetti, Ilia Shumailov, Tianqi Fan, Jamie Hayes, Nicholas Carlini, Daniel Fabian, Christoph Kern, Chongyang Shi, Andreas Terzis, and Florian Tramèr. Defeating Prompt Injections by Design. p. 5 In IEEE Conference on Secure and Trustworthy Machine Learning (SaTML), 2026. p. 7

Linfeng Fan, Ziwei Li, Yuan Tian, Yichen Wang, Rongsheng Li, and Xiong Wang. p. 9 The Granularity Mismatch in Agent Security: Argument-Level Provenance Solves Enforcement and Isolates the LLM Reasoning Bottleneck.

Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. p. 1 Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection. p. 1 In Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security (AISec ’23), pp. p. 2 79–90, 2023.

Keegan Hines, Gary Lopez, Matthew Hall, Federico Zarfati, Yonatan Zunger, and Emre Kiciman. p. 5 Defending Against Indirect Prompt Injection Attacks With Spotlighting. p. 1

David Hofer, Edoardo Debenedetti, and Florian Tramèr. p. 9 Assessing Automated Prompt Injection Attacks in Agentic Environments. p. 8

Feiran Jia, Tong Wu, Xin Qin, and Anna Squicciarini. p. 9 The Task Shield: Enforcing Task Alignment to Defend Against Indirect Prompt Injection in LLM Agents. p. 0 In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. p. 4 29680–29697, 2025. p. 1

Ishan Kavathekar, Hemang Jain, Ameya Rathod, Ponnurangam Kumaraguru, and Tanuja Ganu. p. 5 TAMAS: Benchmarking Adversarial Risks in Multi-Agent LLM Systems, 2025. p. 9

Arseny Kravchenko, Vadim Liventsev, Innokentii Konstantinov, Ildar Iskhakov, and Matvey Kukuy. p. 9 APPA: Recoverable Information-Flow Control for Real-World LLM Agents.

Evan Li, Tushin Mallick, Evan Rose, William Robertson, Alina Oprea, and Cristina Nita-Rotaru. ACE: A Security Architecture for LLM-Integrated App Systems. In Network and Distributed System Security Symposium (NDSS), 2026a. p. 2

Hao Li, Xiaogeng Liu, Hung-Chun Chiu, Dianqi Li, Ning Zhang, and Chaowei Xiao. p. 5 DRIFT: Dynamic Rule-Based Defense with Injection Isolation for Securing LLM Agents. p. 9 In Advances in Neural Information Processing Systems 38 (NeurIPS 2025), 2025. p. 9

Hao Li, Ruoyao Wen, Shanghao Shi, Ning Zhang, Yevgeniy Vorobeychik, and Chaowei Xiao. p. 1 AgentDyn: Are Your Agent Security Defenses Deployable in Real-World Dynamic Environments? p. 1, 2026b.

Junda Lin, Zhaomeng Zhou, Zhi Zheng, Shuochen Liu, Tong Xu, Yong Chen, and Enhong Chen. p. 1 VIGIL: Defending LLM Agents Against Tool-Stream Injection via Verify-Before-Commit. p. 0 In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. p. 4 9764–9785, 2026. p. 0

Xinhang Ma, Taoran Li, Chaowei Xiao, Zhiyuan Yu, Ning Zhang, and Yevgeniy Vorobeychik. p. 5 AutoDojo: Adaptive Black-Box Attacks Reveal the Limits of IPI Defenses and Task-Specification Effects in LLM Agents. p. 9

Xinhang Ma, Chaowei Xiao, William Yeoh, Ning Zhang, and Yevgeniy Vorobeychik. p. 5 ROPE: Routed Origin Policy Enforcement against Indirect Prompt Injection.

Model Context Protocol. p. 1 Model Context Protocol Specification. p. 1

OpenAI. GPT-4o mini: Advancing Cost-Efficient Intelligence. p. 0

Protect AI. p. 5 deberta-v3-base-prompt-injectionv2. p. 0 Hugging Face model card. p. 0

Tianneng Shi, Jingxuan He, Zhun Wang, Hongwei Li, Linyu Wu, Wenbo Guo, and Dawn Song. p. 9 Progent: Securing AI Agents with Privilege Control. p. 1

Peiran Wang, Yang Liu, Yunfei Lu, Yifeng Cai, Hongbo Chen, Qingyou Yang, Jie Zhang, Jue Hong, and Ye Wu. p. 5 AgentArmor: Enforcing Program Analysis on Agent Runtime Trace to Defend Against Prompt Injection. p. 0

Fangzhou Wu, Ethan Cecchetti, and Chaowei Xiao. System-Level Defense against Indirect Prompt Injection Attacks: An Information Flow Control Perspective. p. 1

An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. p. 5 Qwen3 Technical Report. p. 3

12

The paper is 2611.07222. EVERY sentence must end with a reference token. p. 3 p. 3 p. 3 p. 3 The token names the page the sentence came from and quotes the words it rests on, copied word for word from that page>. p. 3

This paper is 2611. p. 0 p. 0 p. 007224. EVERY sentence must end with a reference token. The token names the page the sentence came from and quotes the words it rests on>. p. 3 p. 3

This paper is 2611.0724. EVERY sentence must end with a reference token. The token names the page the sentence came from and quotes the words it rests on>.

This paper is 2611.0724. EVERY sentence must end with a reference token. The token names the page and quotes the words it rests on>. p. 3

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Target admission from typed tool returns with no per-task routing: The system can determine if a proposed side-effecting call is admitted based on a set lookup over identities read from typed tool returns before serialization, ensuring that the floor 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 (Page 4, Algorithm 1).

  2. Sanitization coupled to target admission: The system can remove instruction-like regions its fixed rules match before the model reads them and retains the removed spans as evidence, and this evidence ensures that a local judge may restore a refused call, never one whose target lies in the removed text (Page 4).

  3. A head-to-head comparison with routed-origin enforcement: The system can achieve a zero in-scope attack success on all seven delegated-routing suites and both ChatInject suites, which is at higher utility than routed-origin enforcement on every suite, where it leaks on three and on two (Page 3).

  4. Utility under attack maximization: The system reaches the highest observed mean utility and composite scores (NRP, ERS) of the defenses we run under that attack, holding a mean utility of.484 against an undefended agent under PISmith (Page 7).

  5. Low-latency execution: The defense runs inline with almost no latency and carries to a new tool set once its roles are declared, allowing it to be integrated into many services without significant overhead, as evidenced by adding almost none model invocations over the undefended agent (Page 8).

Sources

Related papers