Paper review · November 18, 2026
SPILLWAY: defending tool-using agents against prompt injection by separating target authorization from input sanitization
SPILLWAY: Separating Target Admission from Sanitization to Defend Tool-Using Agents against Prompt Injection
SPILLWAY deterministically authorizes action targets using only the typed fields returned by tools and removes instruction-like text with rules, blocking all target hijacking by standard injections (zero cases) without extra model calls.
Problem
Agents that handle email, files and payments can be led by instructions hidden in web pages or documents to send data or money to an attacker (indirect prompt injection). Existing defenses use an LLM to make authorization decisions, so that LLM can also be influenced by attack text, and model calls increase at every step. Provenance-based defenses also need a planner or router model for each task.
Method
The method relies on the fact that target values such as recipients and accounts arrive in typed fields of tool returns, while injected instructions arrive in free text such as message bodies. At the output stage (Egress Floor), a target value that is not in the user's task and appears only in tool output is allowed only if it came from a typed identifier field (email, URI, recipient, etc.) of a read-only tool and is not repeated in free text. This decision is a set lookup, so it is the same regardless of the backbone model. Fixed delegation rules (for example, an address containing a name the user mentioned) are also allowed. At the input stage, a deterministic sanitizer removes forged chat templates and instruction-like regions before the model reads them, and keeps the removed spans as evidence. Only rejected calls are re-examined by a local Judge model from two perspectives (routing and payload) and can be restored, but if the target lies inside a removed span, it is not passed to the Judge.
Results
- Across 12 cells in AgentDojo (3 backbones × 4 suites), In-scope attack success for standard template injection was 0 in every cell. The authors reported that it was the only defense among those compared to achieve this.
- Under the same conditions, MELON, Progent and DRIFT were successfully attacked in 7 to 11 of the 12 cells. With qwen3:8b, they added 2.2 to 13.6 backbone calls per benign episode. SPILLWAY did not rerun the backbone for defense.
- Under standard attacks, it had the highest utility under attack and the highest composite scores (NRP, ERS) among the defenses run.
- The raw success rate of the ChatInject (chat template forgery) attack was 0, and the average In-scope attack success rate of PISmith, a public reinforcement learning attacker, was 0.005.
- Across 7 delegation-routing suites, it had higher utility in every suite than ROPE, the closest defense. ROPE was successfully attacked in 3 suites.
Limitations
The guarantee covers only the 'target' of an action. Content attacks, such as sending injected content to a legitimate recipient or injected text mixing into the final answer, are out of scope. Cases where the tool provider writes attacker values into identifier fields are also excluded from the threat model. Tool role classification depends on a fixed dictionary for each benchmark. Evaluation is limited to AgentDojo and AgentDyn, and there is no direct comparison with CaMeL.
Why it matters
The paper proposes a deterministic, auditable rejection criterion for agent security that does not rely on LLM judgment. Because it adds no model calls, teams running email or payment agents can consider it as a baseline defense without extra cost.
Key terms
- Indirect prompt injection
- An attack in which an attacker hides instructions in web pages, documents or emails that an agent reads, in order to steer the agent's actions.
- Taint-novel value
- A value that does not appear in the user's request and appears only in tool output. Because an attacker may have planted it, it is subject to separate verification.
- Egress Floor
- A minimal line of defense that uses rules to check, just before a tool call with external effects, whether the target is in the allowed set.
- Span Veto
- A rule that prevents the Judge model from restoring a target located inside a text span removed by the sanitizer.
Original
Authors: Woosang Lim, Yeseul Chang, Jaewoo Lee · Posted 2026-11-04 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.