acceptodds
Withdrawn from ICLR 2027

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

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

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 check with no per-task plan. The check admits a call when each target was named by the user, directly or through a fixed delegation rule, or returned by a read-only tool in a typed field; 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, SPILLWAY is the only defense we evaluate under which no templated injection redirects an action to an attacker's target, it also holds under chat-template forging, and it requires no defense-triggered backbone re-run while keeping the highest observed mean utility under the basic attack. As personal AI agents take over users' email, files, and payments, this separation gives them a deterministic, auditable guard on where their actions go, one that keeps them useful and does not change with the backbone.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
prompt injection, AI agents, LLM agents, tool-using agents, agent security, runtime defense, information flow control, AI safety, AgentDojo, AgentDyn

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_QXpaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces SPILLWAY which is a runtime defense against prompt injection in tool-using agents. It basically checks whether destinations are authorized based on user instructions and typed tool outputs, then sanitizes potentially malicious content using some predefined rules and then uses an LLM judge to reconsider certain blocked actions. An important restriction is that destinations found in sanitized text cannot be approved by the judge. In the paper the authors evaluate SPILLWAY on AgentDojo and additional AgentDyn suites, along with transferred attacks and using three different models.

Critical strengths
  1. In the paper I like that the main destination check happens outside the LLM. It’s a simple set lookup, so it’s cheap, easy to audit and gives the same answer no matter which backbone is used. Basically you don’t have to trust the model’s judgment for most actions.
Critical weaknesses
  1. I think the main issue is that there’s not much evidence the defense works against attacks it hasn’t already seen. The sanitizer is built to recognize the benchmark injection templates and the main evaluation uses those same templates. As per my understanding the AutoDojo attacks were optimized against other defenses, not SPILLWAY and four of the successful cases were then used to write new rules. The authors actually disclose this which is good, but it means the updated results on that set aren’t really held-out. on the held-out Banking set optimized against DRIFT, SPILLWAY still lets 2 of 144 attacks through and in the other held-out set half the injections are just the template. So, No attack is optimized directly against SPILLWAY. Since the paper main claim is zero attack success, I think this is the biggest open question.
  2. It’s not clear how much the coupling between the sanitizer and the judge (contribution 2) actually helps. In Table 3 removing the span veto doesn’t change any attack outcome, because the veto never gets used. So the evaluation doesn’t really show it helping with safe recovery. There’s also no sanitizer-only baseline. Without it, I find it hard to tell how much of the protection comes from the architecture itself versus just catching and removing known attack patterns.
Questions for authors
  1. If all the rules are frozen, how does SPILLWAY do against attacks optimized directly against it on the cases that haven’t been used before?
  2. Can authors report results for the sanitizer alone and for a strong baseline defense that uses the same sanitizer? I think this would actually help show what the architecture itself adds.
  3. Also, In Algorithm 1 it looks like the delegation rule can let a destination through before the span veto is applied to judge recovery. Do all delegation paths also go through the veto?

Please refer to weakness also.

Reviewer_cDnoOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SPILLWAY, a defense against indirect prompt injection for tool-using agents. It has three parts.

Egress floor. A deterministic check that admits a side-effecting call only when every target value coming from tool output was either named by the user or returned in a typed identity field of a read-only tool (e.g., an email’s sender or a transaction’s recipient), and is not repeated in free text.

Ingress sanitizer. A rule-based component that removes forged chat-template turns, regions matching the benchmark’s injection template, text addressing the agent, and claims that a payee’s account has changed, before the model reads tool output. The removed spans are kept as evidence.

Local judge. A model that may restore a call the floor refused, but is never consulted when the refused target lies inside removed text.

The paper states three structural properties (P1–P3) bounding when an injected identity target can be admitted. On AgentDojo (three backbones × four suites), SPILLWAY has zero attack success on all twelve cells.

Critical strengths

A simple and practical idea for deciding where actions may go. Reading admissible targets from typed tool returns, before serialization, replaces a model judgment with a set lookup. This removes the defense’s dependence on a model that may itself read the injection, and avoids per-task planning or routing.

The experiments are abundant and designed carefully. The paper compares against six defenses across twelve cells, adds three stress tests, and reports an ablation separating floor, sanitizer and judge.

Critical weaknesses

The threat model is flawed, excluding the most natural and popular attacker, and the benchmark cannot test this assumption. P1 assumes the attacker writes only free-text fields. In practice, indirect injection most often arrives in an email the attacker sent, a document the attacker shared, so the attacker’s own identity sits in exactly the typed fields that populate I. An instruction such as “reply to the sender with the file,” which names the attacker by role, is not tested. Zero ASR on AgentDojo says little about this, since their injection also follows the free-text convention.

The sanitizer explicitly removes regions delimited by AgentDojo’s injection template (<INFORMATION> and its “important message” form). However, this threat model assumes the defender already knows the attacker's injection template, which low ASR gives a false-sense of security.

There is no adaptive attack against SPILLWAY itself.

Questions for authors

NA

Reviewer_DkLcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SPILLWAY separates prompt-injection defense into a deterministic target-admission check, a rule-based sanitizer, and a model-based recovery step. The admission check indexes identities from typed read-only tool returns and checks proposed action targets against this index, the user’s request, and fixed delegation rules. The sanitizer removes matched instruction-like text and records the removed spans. A two-view judge may recover refused calls, but cannot approve a target found in those spans.

The evaluation covers three backbones on four AgentDojo suites, a seven-suite comparison with ROPE, and additional attack sets. SPILLWAY reports zero executed attack success on the basic twelve-cell grid, with mean utility under attack of 0.484 versus 0.417 without defense. The paper also reports ablations, inference overhead, and residual failures on held-out optimizer-generated attacks.

Critical strengths

The design provides a simple, auditable execution check with limited additional model work. Reading identity fields before serialization avoids asking a model to reconstruct provenance from text. The stricter treatment of grants and destructive operations is also sensible. Reserving the judge for refusals offers a practical way to recover some utility without rerunning the backbone for every decision.

The evaluation reports both security and utility, with useful qualifications. The basic grid spans three backbones, and the ROPE comparison includes tasks that require delegated routing. The paper distinguishes executed effects from blocked calls left in transcripts, reports raw and scoped attack success, and discloses both rule development on failed examples and held-out failures. This makes the empirical evidence easier to assess.

Critical weaknesses

The strongest security results rely on attack-aware sanitization without a direct adaptive evaluation. The sanitizer explicitly recognizes AgentDojo’s injection delimiters and forged chat-template markers. Additional address and payee-change rules were written using four failures from the reported AutoDojo set. The authors correctly identify that result as descriptive rather than held out, but the remaining tests still transfer attacks optimized against other defenses. On the held-out DRIFT-optimized Banking set, the paper reports two attacker-directed transfers out of 144 pairs when both are counted. The current evidence therefore supports strong performance against specific attack families, rather than a robust barrier against an attacker targeting the sanitizer and recovery judge together.

The evaluation does not isolate the benefit of the proposed coupling. Table 3 has no sanitizer-only arm, and removing the span veto changes no attack outcome because the agent never proposes a removed target. Thus, a central architectural feature is not exercised by the main ablation. This matters because the admission index is a flat set: it establishes that an identity appeared in a typed return, not that the identity is authorized for the current action. Attacks referring indirectly to an already indexed identity could leave more responsibility with the sanitizer and judge than the headline deterministic framing suggests. Appendix C explicitly leaves such role-based references untested.

Questions for authors

Assessment-changing comparison: How does the sanitizer alone perform on the basic, ChatInject, and held-out optimizer attack sets? This would establish how much security and utility the admission layer adds beyond removing recognizable attack text.

Assessment-changing experiment: Can you run an attacker with feedback against the complete SPILLWAY pipeline, with the sanitizer rules frozen beforehand? Please include attacks that paraphrase instructions, derive target strings, and refer indirectly to identities already present in the index.

Clarification: Can you provide an evaluated case where the span veto blocks an unsafe call that would otherwise pass both judge views? Its structural restriction is clear, but the current ablation does not demonstrate an empirical benefit.

Clarification: What benign tasks are rejected by the rule that refuses an indexed identity when it also appears in free text? Email signatures, quoted addresses, and document contact details seem likely to trigger this condition. A breakdown would help explain the Slack utility loss and the judge’s recovery workload.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.