acceptodds
Withdrawn from ICLR 2027

Schema Laundering: Embodied Safety Gates — Audit the Rule and Trust the Rewriter

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Embodied agents act on the world, so a command that slips past screening does harm no later correction undoes. The common defence is a typed gate: a language model rewrites each command into fixed fields and a table decides, so the decision is auditable. Such gates are reported against undefended agents, never against their own policy read as prose, which confounds structure with policy. Holding the policy fixed and varying only its form, a single prose reading of the gate's own policy is better by more than the pre-declared margin and buys no measurable precision; the loss concentrates in euphemised rewrites. A typed gate decides not by its rules but by when the rewriter chooses a field, so a paraphrased command is filed as harmless before any rule runs. We call this schema laundering: the protection lives in the policy, not the structure. We build \telos, a training-free command gate under one rule: any signal adds a block, only perceived evidence releases one, and an unseen named object routes to a third outcome, \observe. Read safety-first (a hazard-recall floor), it matches a fine-tuned guard's recall at lower over-block and holds that over-block off the training distribution, where the fine-tune blocks – of benign tasks. Its edge is an unforgeable release that lifts only on perceived normal-use evidence an attacker cannot fabricate, giving the lowest adaptive-attack success; the residual leak is in the harm judge, not the rule. As embodied and agentic systems spread, a training-free gate that stays robust under adaptive rewording and ports across scenes is a practical route to auditable physical-action safety, and a typed safety layer should be audited against its own policy in prose before its structure is credited.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
embodied agent security, robot command screening, LLM guardrails, jailbreak, euphemism evasion, typed representations, schema laundering, information-matched control, benchmark integrity, pre-registration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Fk1pOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies the safety implications of converting natural-language robot commands into structured representations before applying safety rules. The authors argue that this conversion can lose safety-relevant information, allowing euphemistically rewritten harmful commands to evade detection, a phenomenon they call schema laundering.

Critical strengths

A potentially useful empirical observation. The comparison between prose-based policy evaluation and typed policy evaluation raises an important question: whether the benefits of structured safety checks come from the structure itself or from the underlying policy. The reported difference on euphemistically rewritten commands motivates further investigation.

Broad experimental coverage. The paper evaluates multiple safety benchmarks, includes comparisons with both rule-based and learned defenses, and examines adaptive rewriting, ablations, and error cases. The attempt to distinguish in-domain performance from out-of-domain behavior is also useful.

Critical weaknesses
  1. The methodological novelty of TELOS appears limited relative to its claims

TELOS assembles several well‑established building blocks: conservative union‑based composition of safety signals, deterministic physical‑hazard invariant checks, LLM‑driven harm assessment, perception‑based conditions, and selective defer‑to‑observe outcomes for failed grounding. While these individual components are reasonable, the paper does not sufficiently articulate what constitutes its core methodological advance over prior embodied‑agent guardrails and selective‑classification literature. Specifically, it remains ambiguous whether the contribution lies in individual modules, or in the strict one‑way compose‑release discipline (blocks can only be added; only perceived evidence can lift a block, never textual prompting). This conceptual distinction needs clearer elaboration and positioning against prior work.

  1. The experimental results do not fully support the broad claims about generalization and superiority

The paper reports competitive performance across several datasets, but the results are more nuanced than the abstract and conclusion suggest.

  1. The evidence for robustness against adaptive attacks is limited

The adaptive evaluation uses 100 command pairs and an attacker that rewrites refused commands for up to five rounds. TELOS achieves an attack success rate of 5/100, compared with 7/100 for the policy-only baseline.

  1. Writing and presentation create unnecessary barriers for readers.
Questions for authors

see the Weaknesses

Reviewer_EPYQOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies schema laundering in embodied safety gates: a command rewriter can map a euphemised harmful request into apparently benign typed fields before a rule checker runs. It compares a typed implementation with a prose reading of the same policy, reporting 294/400 versus 328/400 conceptual-deception attacks blocked at equal benign over-block. It then proposes TELOS, a training-free gate that combines policy, invariant, and harm-judge signals, conditionally releases policy-only blocks using perceived normal-use evidence, and defers unseen-object cases to OBSERVE. Decision-layer evaluations include two SafeAgentBench views, household and situational datasets, and a five-round adaptive attack suite.

Critical strengths
  1. A consequential control for embodied safety evaluation. Separating policy content from a particular typed implementation addresses a real confound in evaluating rule-based safety layers. The paired-command comparison, the reported laundering signature in 37/61 discordant cases, and the second model-family result provide useful evidence that semantic rewriting deserves explicit auditing.

  2. Useful attention to both safety and usability. The paper reruns several guards on matched items, distinguishes image and simulator-state views, and explicitly acknowledges that the full system does not reduce binary over-block over Policy-only on SAB567, performs poorly on HomeSafe hazards, and has no real-robot evaluation. The OBSERVE outcome also offers a reasonable way to distinguish insufficient evidence from a definitive refusal.

Critical weaknesses
  1. The principal mechanism and module contributions are not sufficiently isolated. Table 1 labels the prose arm as the full TELOS gate, although the text describes a single Policy Reader; their equivalence needs to be established. More decisively, Section 6.6 calls Judge-authority a removal of the Release Gate, but Section 4.6 defines it as also removing the Policy Reader's blocking authority. With fixed policy/invariant/judge outputs, disabling a block-release operation cannot lower non-ALLOW hazard recall. The reported 94.1% to 63.1% change therefore does not isolate Release Gate. The observation supports a weakness of the evaluated typed pipeline, but not the general conclusion that structure contributes nothing or that immutable perception makes the overall gate impossible to bypass.

  2. Central effectiveness and generalisation claims need consistent accounting and stronger qualification. Section 6.7 says reconsideration never changes a block to ALLOW, yet reports recovering 140 over-blocks; BLOCK-to-OBSERVE recovery would not correct binary over-block. The stated net +158 also needs the complete additional-error accounting, since 140 + 21 = 161 before such costs. Table 4 calls HomeSafe784 HomeGuard's training set, while HomeGuard Section 4.1 explicitly identifies the matching 512-unsafe/272-safe HomeSafe-Bench as non-overlapping with training. Moreover, TELOS recalls only 53.7% of HomeSafe hazards, and MSS hazard recall is not reported alongside its low over-block. Finally, 5/100 versus 7/100 adaptive successes is too small to establish an advantage over the strongest Policy-only baseline, and HomeGuard is absent from that comparison. These gaps affect the interpretation of the system's main benefit.

Questions for authors
  1. What is the pre-declared margin, and does resampling preserve the original command pairs and attack variants?

  2. Can you report an ablation that only disables Release Gate with all policy/invariant/judge outputs fixed? Please also define the normal-use predicate, action-object binding, and priorities among BLOCK, ALLOW, and OBSERVE.

  3. In Section 6.7, do recovered over-blocks become ALLOW or OBSERVE? .

  4. Is HomeSafe784 the held-out HomeSafe-Bench identified in HomeGuard Section 4.1?.

  5. Please provide the paired adaptive outcomes and uncertainty for TELOS versus Policy-only, clarify how harmful-intent preservation is checked, and either include the matched HomeGuard comparison or limit the robustness claim to the tested competitors. These are clarification or validation requests that could change my recommendation.

Reviewer_UdJdOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies whether a typed safety gate can lose dangerous semantics when a language model rewrites a natural-language command into fixed fields before a rule system evaluates it. The authors call this failure mode schema laundering. They compare prose and typed forms of the same policy on RoboJailBench, then propose TELOS, a training-free decision-layer gate combining a Policy Reader, Invariant Bank, Harm Judge, Composer, perceived-evidence Release Gate, and Re-perception Probe with an OBSERVE outcome.

Critical strengths
  • The problem is important and well motivated. Auditability of fixed rules does not necessarily imply semantic safety.

  • The prose-versus-typed comparison, TELOS ablations, OBSERVE mechanism, and off-domain over-block results provide useful initial evidence.

Critical weaknesses

The current results show a performance difference between two complete pipelines, but do not yet establish that the difference is caused solely by the representation format or that protection comes from the policy rather than the structure.

Questions for authors
  • The paper notes a similar format-over-content effect in tool specifications, but does not define when this phenomenon should be called “schema laundering” rather than ordinary semantic compression or information loss.
  • The threat model assumes that the attacker cannot alter the perceived scene. The evaluation therefore does not cover sensor spoofing, scene manipulation, object replacement, occlusion, perception errors, stale states, timing gaps between re-perception and execution, action failures, or actuator faults.
  • The adaptive attack results, 5/100 versus 7/100, may reflect sampling variability. Without uncertainty estimates or broader attacker evaluations, they do not support a general claim of adaptive robustness.
  • The paper reports one bootstrap interval and one p-value, but does not consistently provide uncertainty or paired tests.
  • The authors acknowledge that some SafeAgentBench unsafe labels are debatable. Label noise may affect the reported Harm Judge ceiling and could lead to incorrect attribution of errors to the judge rather than to the rule, grounding, or release mechanism.
  • The paper lacks representative item-level examples and sufficient reproducibility details, including the exact policy, schema, prompts, invariants, release predicates, thresholds, parsing rules, and module outputs.

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

Related papers

Discussion (0)

Sign in to comment.