acceptodds
Under review as a conference paper at ICLR 2027

When the World Lies: Backdoor Attacks on Latent World Models for Downstream Control

Abstract

Pretrained world models, learned simulators that encode an observation into a latent state and predict how it evolves under actions, are beginning to be reused as off-the-shelf dynamics backbones for control. We study a supply-chain backdoor in which the attacker controls only the released checkpoint, while the victim trains on clean data and never sees the trigger during training. To our knowledge, this is the first checkpoint-only world-model backdoor that induces targeted actions through a downstream controller trained or instantiated using clean data. The attack encodes no explicit trigger-to-action rule. Instead, the poisoned model routes trigger-bearing observations into a chosen latent region and reshapes the local dynamics there, so that the victim's own optimization, through Dreamer-style actor training or MPC/CEM planning, re-discovers the attacker's target action on its own. Across several control tasks and trigger families, the strongest evaluated settings control every action dimension and reach up to 100% Step-ASR, while harder and transfer settings show partial target recovery. Across the main settings, poisoned checkpoints retain roughly 74-100% of baseline clean utility and can evade clean-data diagnostics in most evaluated cases. Matched clean-model controls separate the backdoor from the effect of the trigger alone, and mid-episode tests show that the action changes only while the trigger is present. Moderate trigger-blind repair can leave triggered failure intact, while stronger adaptation trades off against clean control. These results identify the released world-model checkpoint as an underexamined attack surface for downstream control.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Backdoor attacks, World models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9zp9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a backdoor attack against pretrained World Models. The attacker only needs to modify the World Model checkpoint before its release, without interfering with the victim's training data, reward function, or controller. Even if the victim trains a new Actor using clean data or directly uses the World Model for MPC/CEM planning, the controller may still execute an attacker-chosen action when a specific trigger appears.

The authors propose two attacks for these two control settings: Latent-Trap Attack and CEM-Plan Attack. The former modifies the latent dynamics to make a newly trained Actor prefer the target action. The latter changes the costs predicted by the World Model so that the MPC/CEM planner selects the attacker's desired action. The authors conduct experiments on DreamerV3 and LeWorldModel to evaluate attack effectiveness, different trigger types, and the impact of the backdoor on normal model performance and defense methods.

The paper addresses a meaningful security problem. As pretrained World Models are increasingly reused in downstream systems, understanding whether these models can be trusted is an important research question.

Critical strengths

S1. The research problem is meaningful, and the attack setting has some novelty.

This paper shows that an attacker can influence a victim's downstream controller by modifying only the World Model checkpoint. In particular, the backdoor can still work even when the victim trains a new Actor using completely clean data. This reveals a security risk when pretrained World Models are reused in downstream tasks, which I consider the paper's most valuable contribution.

S2. The experiments provide convincing evidence of attack effectiveness and validate the attack mechanisms.

The authors demonstrate effective attacks in both Actor and Planner settings. Through comparisons with clean models, experiments that turn the trigger on and off, and component ablations, they show that the attacks are caused by the poisoned World Model rather than ordinary visual perturbations or general performance degradation. These experiments make the paper's conclusions more convincing.

Critical weaknesses

W1. The attack may require prior knowledge of downstream tasks to select suitable target actions.

Although the authors conduct some generalization experiments, the results seem to suggest that the attack works better when the target action is consistent with the normal behavior of the downstream task. This means that attackers may need to know what actions the downstream task typically performs in order to choose an effective target action. If such detailed knowledge of downstream behavior is required, the practical applicability of the attack may be limited.

W2. The attack's dependence on the downstream controller mechanism may affect its stealthiness.

The attacker needs to know the type of downstream controller in advance and design different attacks for different control mechanisms. If the victim uses another control mechanism, the attack may fail. More importantly, since the backdoor modifies the internal dynamics of the World Model, it may still reduce normal task performance even when it fails to induce the target action. This could make the backdoor easier to detect. The paper currently lacks experiments examining this possibility.

W3. Performance drops on most normal tasks weaken the evidence for the backdoor's stealthiness.

As shown in Table 1 and Table 2, normal task performance decreases after backdoor injection in all tasks except Walker walk. Some drops are relatively large. For example, the return on Quadruped walk decreases from 697.4 to 519.1. It remains unclear whether these drops exceed the normal variation between model training runs and whether the victim could detect the backdoor through normal task evaluations.

Questions for authors

C1. Attack effectiveness under different control mechanisms

I suggest that the authors explain or provide additional experiments to examine whether the same backdoored World Model could still reduce normal task performance when the victim uses an unexpected control mechanism, even if the attack fails to induce the target action.

C2. Normal performance degradation and stealthiness

I suggest that the authors compare the performance drops in Table 1 and Table 2 with normal model performance variations, and analyze whether these drops could allow the victim to detect the backdoor through normal task evaluations.

Q1. How are target actions selected?

Does the attacker need prior knowledge of the downstream task's normal behavior to select an effective target action? How do the authors select target actions in practice?

Q2. Differences from TrojanWorld

TrojanWorld also studies how poisoning a World Model can influence the actions of downstream agents. What are the key differences between this paper and TrojanWorld in terms of their threat models and technical contributions?

2 comments

We thank the reviewer for the thoughtful discussion and for identifying the clean-data training of a fresh downstream actor as a central contribution. We appreciate the positive assessment of the mechanism evidence and the controls separating targeted steering from ordinary visual disruption.

W1/Q1 Target actions and task knowledge. We thank the reviewer for raising this point. We chose the targets manually within the valid action space to make the induced behavior easy to recognize. For Walker, we used an alternating action pattern on both legs. These targets were not tuned separately to each downstream objective, and poisoning does not use the victim's future reward or goal.

Appendix B discusses how task compatibility can make some targets easier to enforce. However, the full Walker target occurs in only 0.21% of clean walking steps, yet the attack recovers all six joints. We also keep the same poisoned model and target when changing the victim's objective from walking to running.

In the running task, four of six joints remain controlled, while two consistently change sign. The actor therefore produces a different, repeatable action combination, while the trigger still activates the attack. Task compatibility can affect how precisely the original target is recovered, but incomplete recovery does not necessarily remove the triggered steering or task failure.

W2/C1 What happens with another controller? Following the reviewer's suggestion, we replaced the Walker actor with a CEM planner on the same supplied models. The poisoned models retain useful clean control, with mean return 853.0, but return falls to 24.2 when the trigger is added. The benign controls remain stable. Target cosine is only 0.297 for the poisoned models, so precise targeting does not transfer, but the triggered disruption remains. Changing the controller therefore leaves useful clean control alongside severe temporal trigger-gated disruption in this setting.

Walker actor-to-planner transfer (n=5 victim runs; 20 episodes/run).

ModelClean returnTrig. returnTrig. cos.
Benign747.7 ± 99.9734.5 ± 88.0-0.171 ± 0.035
Poisoned853.0 ± 26.324.2 ± 5.50.297 ± 0.121

W3/C2 Clean performance and training variation. We now report five downstream runs per condition to compare the clean-performance gaps with training variation. The effect depends on the task: Cheetah loses about 2% of clean return, whereas Quadruped can lose about 25%, which is larger than the observed variation between runs. We therefore cannot attribute the Quadruped drop to ordinary seed variation. A victim with a trusted performance reference could notice this loss. However, a recipient evaluating a downloaded model without such a reference would not have the same comparison.

Clean returns (n=5 downstream victim runs/group).

TaskBenignPoisonedDrop (%)
Quad. walk691.2 ± 24.7519.2 ± 13.924.9
Cheetah run877.2 ± 7.0859.8 ± 2.82.0

Q2 Difference from TrojanWorld. We thank the reviewer for pointing out this related work. Both papers study backdoors implanted through the world model. The main difference is when the downstream actor is trained. TrojanWorld uses an already trained, frozen actor or decision module as feedback during poisoning. In our actor setting, the victim trains a new actor, critic, and reward head on clean data after receiving the poisoned checkpoint. Our latent-trap construction makes this new actor learn the malicious preference through imagined rollouts. For the planner setting, we differentiate through a softCEM surrogate to influence the iterative search. TrojanWorld additionally focuses on persistence after trigger removal, while we emphasize the trigger-gated behavior as it gives the attacker more control over when the backdoor is active or hidden. We will add TrojanWorld to the related work and to the comparison in Table 18.

Ethics. We also acknowledge the reviewer's concern about releasing attack code and poisoned models. We will discuss the risks of these artifacts alongside the detection and mitigation results. To limit misuse, we could keep specific poisoned checkpoints and parts of the attack code private, making them available only upon request for reproducibility and defense research.

Thank you for the detailed response and additional experiments. The new results on controller transfer and clean-performance variation help address my concerns and clarify both the effectiveness and limitations of the proposed attacks.

I particularly appreciate the authors' acknowledgment that the clean-performance degradation on Quadruped exceeds normal training variation and may affect the attack's stealthiness. While this concern is not fully resolved, I believe the additional evidence and clarifications have sufficiently addressed my main questions. I am willing to raise my rating from 2 to 3.

Reviewer_MTG3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies backdoor attacks on pretrained world models, where the attacker supplies the model weights that a victim uses for downstream tasks. During deployment, an observation trigger routes latent states toward an attacker-chosen anchor, inducing the controller, which can either be policy or planner, to favor a target action. Empirical results demonstrate that the target action can be recovered via the observation trigger in several settings.

Critical strengths

The paper considers a novel attack scenarios, where the backdoor is injected to the world model to induce a target action without compromising the victims' performance. It also considers 2 complementary settings, i.e., model-based RL and planning, and show how attacks can be injected can steer the controllers to choose the target action.

Critical weaknesses
  • For the dreamer-style actor, the attack relies on the actor encountering anchor states during imagined rollouts. However, these rollouts may reveal that the actor is trapped into some low-value absorbing states. This may serve as a diagnostic signal to help the victim to identity a compromised world model.
  • A victim could learn a residual dynamics model from real environments data to correct errors in the supplied world model. Such adaptation might weaken the attack.
  • The conditions for a successful attack appear restrictive, as the attacker must both modify the model weights and manipulate the victim’s observations at deployment.
Questions for authors

Please refer to the weaknesses.

2 comments

We especially thank the reviewer for the constructive suggestions on the detector and residual model. We also appreciate the reviewer's recognition of the checkpoint-reuse threat and the complementary actor/planner settings.

W1 Diagnostics and adaptive detection. We first checked whether the trap could be found by imagining rollouts from clean observations. For each task, we evaluated 192 rollouts per group (two independent models, 16 clean episodes per model and two different starting points per episode with three seeds each). We assume that the state is a trap when the predicted reward is unusually low over the last 20 steps, the latent state changes much less than at the start, and trying different action sequences from that state produces little difference. On three of the four tasks, benign and poisoned models produced the same number of flags. Walker run was the exception, with 9/192 flags for poisoned models and 0/192 for benign models. This result comes directly from the attack design. The trap only reshapes transitions that start at the anchor, a specific point in the latent space, so a clean rollout must reach z⋆z^\star and take a non-target action several times before it becomes obvious that this specific latent region has been modified to become a trap.

Clean imagined absorbing-event flags.

TaskBenign flagsPoisoned flags
Walker walk0/1920/192
Cheetah run0/1920/192
Quadruped walk1/1921/192
Walker run0/1929/192

We therefore turned the reviewer's idea into an active detector. Instead of waiting for rollouts to reach the trap, it searches for inputs that collapse the posterior onto a single latent. This posterior-collapse score separates poisoned from benign victims on all four tasks (AUROC 1.000, n=5 per group), while an encoder-only score is inconsistent across tasks. The same patch search also flags all four poisoned trigger families in the planner setting (see our response to Reviewer 21Lw).

Actor detector (n=5 related victims/group; descriptive AUROC).

TaskEnc. AUCClean post. scorePoison post. scorePost. AUC
Walker walk0.7201.329 ± 0.0707.807 ± 0.0291.000
Walker run1.0001.277 ± 0.0978.249 ± 0.0151.000
Cheetah run1.0000.580 ± 0.01610.292 ± 0.0101.000
Quad. walk0.0001.254 ± 0.0427.711 ± 0.0461.000

The detector needs no trigger, target action, or anchor, and sets its threshold on benign reference models, as checkpoint-level trojan detectors do. We will include it in the revised paper as a possible adaptive defense against our attack.

W2 Residual correction. The reviewer's intuition is supported by the planner experiment. A residual fitted to clean real transitions reduces target cosine from 0.962 to 0.203, while clean task success changes from 84% to 74%. However, triggered task success remains low, maintaining the trigger-gated failure.

Planner residual correction

Model / correctionCA (%)Trig. SR (%)Trig. cos.
Poisoned / none84.0 ± 4.013.0 ± 1.00.962 ± 0.016
Poisoned / small64.0 ± 4.08.0 ± 2.00.096 ± 0.001
Poisoned / large74.0 ± 4.09.0 ± 3.00.203 ± 0.000
Benign / none78.042.00.036
Benign / small59.0 ± 5.030.0 ± 6.00.008 ± 0.010
Benign / large85.0 ± 1.035.0 ± 7.00.009 ± 0.006

W3 Threat model. We thank the reviewer for raising this point. Our threat model follows the established model-supply-chain backdoor setting: the attacker provides a poisoned checkpoint, and the backdoor activates when the chosen trigger appears in the victim's observations. Introducing the trigger at deployment is a standard part of this setting. The attacker does not need to modify the victim's downstream controller or training objective.

I thank the authors for their clarifications and additional experiments, which address most of my concerns. I have therefore raised my score from 2 to 3. However, given my background, I remain uncertain about the broader significance of this work for AI security and its practical implications.

Reviewer_21LwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies checkpoint-only backdoor attacks against pretrained latent world models which are used for downstream control. The attacker modifies only the world-model checkpoint, while the victim trains an actor or deploys a planner using clean data, without exposure to the trigger during training. The paper proposes two controller-aware attacks: a latent-trap attack targeting Dreamer-style policy learning and a planner-facing attack using a differentiable softCEM surrogate for MPC/CEM planning. Experiments show strong triggered degradation and target-action recovery while preserving reasonable clean performance in some settings.

Critical strengths

This paper investigates an underexplored checkpoint-only world-model backdoor that induces targeted actions through a downstream controller trained or instantiated using clean data. The actor-facing attack creates a target-action fixed point and an absorbing non-target region, whereas the planner-facing attack shapes the cost surface through a differentiable approximation of iterative CEM optimization. These two controller types are technically sensible. The evaluation covers both actor-based and planning-based control, including multiple environments, different trigger families, and controller hyperparameter variations.

Critical weaknesses

W1: This paper characterizes both specific controller attacks. But in a real supply-chain scenario, the attacker doesn’t know whether the victim will use Dreamer-style actor training or MPC/CEM planning. The paper does not evaluate cross-controller transfer that whether an actor-poisoned checkpoint remains effective when used by a planner.

W2: The paper evaluates several trigger types but does not assess their robustness under realistic sensing conditions. The feasibility of reliably activating these triggers in physical environments is important to establishing the practical threat.

W3: Step-ASR is defined using a cosine similarity threshold of 0.8. This metric measure is not enough to evaluate the exact reconstruction of the target motion vector. Two motions may have very high cosine similarity but still exhibit significant differences in amplitude.

Questions for authors
  1. The paper assumes controller-specific attacks, but the attacker may not know the downstream controller in a realistic supply-chain scenario. Can the authors evaluate cross-controller transfer or develop an adaptive attack that remains effective across both Dreamer-style actors and MPC/CEM planners?
  2. Given the small number of independently trained checkpoints, how reliable is the claim that poisoned models evade clean-data diagnostics? Can the authors report detection rates at fixed false-positive rates and evaluate a broader range of adaptive, checkpoint-level detectors?
  3. Can the authors provide an attack-strength versus clean-utility analysis? How sensitive are the results to the poisoning budget and the weights of the attack and clean-preservation losses?
  4. How much of the observed task degradation is attributable specifically to the backdoor, rather than to the trigger's out-of-distribution effect?
2 comments

Part 1

We thank the reviewer for the careful assessment and for recognizing the importance of the checkpoint-only threat and the two complementary controller settings. We address the requested controls below.

W1/Q1 Controller transfer. Even though the attacker cannot be sure how the model will eventually be used, they can make an informed choice based on the model they are poisoning. Dreamer is designed to train an actor using imagined rollouts, while LeWorldModel is used for planning. The attacker could also suggest this setup when publishing the pretrained model, although the victim might end up using a different controller.

To check what happens in this case and following the reviewer's suggestion, we replaced the Walker actor with a CEM planner at five victim checkpoints using the same originally poisoned world model, running 20 episodes each. Triggered return falls from 853.0 to 24.2 on poisoned models, while benign return remains stable. Thus, the trigger-gated task disruption transfers, even if accurate targeting does not.

Walker actor-to-planner transfer (n=5 victim runs; 20 episodes/run).

ModelClean returnTrig. returnTrig. cos.
Benign747.7 ± 99.9734.5 ± 88.0-0.171 ± 0.035
Poisoned853.0 ± 26.324.2 ± 5.50.297 ± 0.121

W2 Robustness. We agree that it is important to check whether the trigger still works under changes in sensing conditions. We tested all four trigger families under blur, sensor noise, and brightness changes, and included benign-model controls. For the patch trigger, we also compared changes applied to the whole image with changes applied only to the patch. Target alignment remains high under most perturbations, even when clean task success drops. Changes limited to the patch also have little effect on targeting in most tested conditions. Different triggers show different weaknesses: body color is more sensitive to changes in brightness, while the additive pattern is more affected by blur. These results suggest that robustness depends on the chosen trigger, rather than being a fixed property of the attack.

Sensing robustness: Clean SR (%) / target cosine (20 episodes).

TriggerRef.Blur σ=3Noise σ=0.06Light ×0.5Light ×1.25
Patch75 / 0.97040 / 0.97080 / 0.9705 / 0.9495 / 0.971
Body color65 / 0.97015 / 0.97055 / 0.97010 / 0.46215 / 0.236
Color cast80 / 0.97120 / 0.97270 / 0.9715 / 0.1085 / 0.971
Additive (32/255)85 / 0.97225 / 0.45565 / 0.97210 / .1705 / 0.970
Benign90 / 0.01175 / 0.00985 / 0.05020 / 0.12515 / 0.185

Patch-only versus whole-image corruption: CA (%) / target cosine.

Model/scopeRef.Blur σ=3Noise σ=0.06Light ×0.5Light ×1.25
Benign / whole90 / 0.01175 / 0.00985 / 0.05020 / 0.12515 / 0.185
Benign / patch85 / 0.03085 / 0.03390 / 0.02575 / .00575 / 0.009
Poisoned / whole75 / 0.97040 / 0.97080 / 0.9705 / 0.9495 / 0.971
Poisoned / patch75 / 0.97075 / 0.97080 / 0.97075 / 0.97180 / 0.970

W3 Direction and magnitude. We appreciate this distinction. We agree that cosine alone does not measure action amplitude. Appendix A, Tables 4-5 also compare each joint's mean action with its target value, checking both its sign and amplitude. On Walker walk and Cheetah run, all six joints match the target amplitude to the reported precision. On Quadruped, all 12 joints have the target sign, and eight have mean amplitudes within 0.2 of the target, the remaining four falling a bit short of the limit. We will make this clearer in the paper and refer to the result as target-directed action control, since the controller follows the chosen direction even when it does not fully match the target amplitude.

Part 2

Q2 Detection.

Clean-data diagnostics do not separate poisoned checkpoints, but mechanism-aware detection does. We distinguish two settings. The checks a recipient normally runs before deployment (reconstruction error, one-step prediction error, posterior--prior KL, and clean task return) do not expose the backdoor: three of four poisoned checkpoints fall inside the benign distribution on every metric (Appendix E).

Following the reviewer's suggestion, we then built two checkpoint-level detectors that target our attack mechanism directly. The rollout-variance detector looks for unusually low latent variance in imagined rollouts from clean observations. The patch-inversion detector searches for a patch that maps different clean images to a common latent, which is the effect the encoder teleport installs. Neither detector uses the trigger, the target action, or the anchor.

We set both thresholds to a 5% false-positive rate and calibrated them on 20 benign models trained from the same pretrained checkpoint with matched clean training settings. We then tested them on 20 separate benign models and 10 poisoned models. Both detectors flag all 10 poisoned models, patch inversion stays at a false-positive rate of (1/20), while rollout variance is noisier (3/20). The same patch search also flags all four trigger families (5/5 each), because every family is trained to induce the same latent collapse.

Checkpoint detection using thresholds calibrated on 20 trusted benign models at a nominal 5% false-positive rate.

DetectorFalse positivesPoisoned detected
Rollout variance3/20 (15%)10/10 (100%)
Patch inversion1/20 (5%)10/10 (100%)

Identical patch-inversion detector across trigger families. Scores are mean ± population standard deviation over five model runs per group. Higher scores indicate stronger induced latent collapse; the detection threshold is 3.4262.

Model / original triggerInversion score ↑Flagged
Benign3.382 ± 0.0210/5
Patch9.232 ± 0.1245/5
Body color6.356 ± 0.0425/5
Global color cast8.618 ± 0.2595/5
Additive pattern (32/255)6.834 ± 1.1375/5

We will add both detectors to the revised paper as a possible defense contribution and state the stealth claim accordingly: the attack evades the clean-data diagnostics a recipient would normally run, while an adaptive defender with knowledge of the attack mechanism can detect it.

Q3 Attack strength and clean utility. We varied the attack weight and poisoning duration on Reacher. Increasing the attack weight does not necessarily improve the attack. At ×4, clean success falls from 84.7% to 66.7%, while target cosine also decreases from 0.943 to 0.878. Giving the attack loss too much weight can therefore harm clean performance without improving targeting.

The benefit of stabilization is also shown in Appendix D: clean success increases from 68% to 80% on Reacher, and from 36% to 84% on TwoRoom. However, increasing preservation further is not always beneficial. In the new Reacher comparison, doubling the clean-teacher weight reduces clean success from 84% to 75%, while target cosine remains approximately 0.962. Stronger preservation may restrict the model's ability to adapt during training, without necessarily weakening the triggered behavior. Longer poisoning also helps in the tested baseline runs. From epoch 1 to epoch 4, target cosine increases from 0.847 to 0.962, and clean success increases from 76% to 84%.

Attack-weight sweep on Reacher after four epochs. n denotes the number of training seeds.

Attack weightClean SR (%)Trig. SR (%)Target cosine
×0.582.0 ± 0.08.0±2.00.949 ± 0.015
Baseline84.7 ± 3.411.3 ± 2.50.943 ± 0.030
×275.0 ± 3.010.0 ± 2.00.935 ± 0.012
×466.7 ± 2.58.0 ± 2.80.878 ± 0.024

Poisoning duration on Reacher. Mean ± SD across the same two training seeds, evaluated at fixed checkpoints within a four-epoch schedule.

EpochUpdatesClean SR (%)Trig. SR (%)Target cosine
1200076.0 ± 0.05.0 ± 1.00.847 ± 0.039
2400074.0 ± 2.05.0 ± 1.00.916 ± 0.014
4800084.0 ± 4.013.0 ± 1.00.962 ± 0.016

Q4 Trigger OOD effects. Thank you for raising this point. As shown in Appendix J (Table 17), the same trigger has little effect on clean-model performance in most tasks. Reacher is more sensitive, but the trigger still does not produce target-directed actions or the same strength of success-rate drop. These controls show that the failures on poisoned models cannot be explained by the trigger's OOD effect alone.

No-trigger → triggered performance. Clean controls from Table 17; poisoned means from Tables 1-2.

TaskClean WMPoisoned WM
Walker walk (return)940→940961.3→41.9
Cheetah run (return)887→875860.6→0.7
PushT (success)96%→94%82.0%→1.3%
Reacher (success)81%→52%78.0%→10.6%

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and additional experiments

We thank the reviewers for their comments and suggestions, which have helped us strengthen the evaluation and clarify the scope of the paper. We have revised the manuscript to incorporate the additional experiments and clarifications provided during the discussion. Changes are highlighted in blue. Below, we summarize the revisions addressing each reviewer’s comments:


Reviewer 21Lw,

We incorporated the actor-to-planner transfer experiment, showing that useful clean control is retained while triggered disruption survives the controller change. We also evaluated all four trigger families under simulated blur, sensor noise, and changes in brightness. Triggered targeting remains strong under most tested perturbations, including conditions that substantially reduce clean task success, while changes limited to the patch have little effect on targeting. We also clarified the per-joint measurements of action sign and amplitude. We implemented and evaluated calibrated rollout-variance and patch-inversion detectors, as well as posterior-based detection, in the actor setting. Both calibrated detectors flag all ten evaluated poisoned models, we present them as possible detectors against our attack when clean baselines are available. Finally, we added attack-strength and clean-preservation comparisons and clarified the existing robustness results across planner configurations and actor imagination horizons.


Reviewer MTG3,

We incorporated the additional clean-start rollout diagnostics, posterior-based detection, and clean-data residual correction experiments. Residual correction substantially weakens planner targeting, while triggered task success remains low under the tested correction budgets. We appreciate the constructive suggestions and follow-up feedback, which helped strengthen the defense evaluation.


Reviewer 9zp9,

We incorporated the controller-transfer experiment and clean-performance comparisons across five downstream runs per condition. We explicitly distinguish the larger Quadruped performance reduction from the smaller changes in the other actor settings. We also clarified target-action selection and the comparison with TrojanWorld, and updated the ethics and artifact-release statements. We appreciate the thoughtful comments and follow-up feedback.


The new experiments are discussed in Section 5 and Appendix L, with the existing controller-configuration evaluations in Appendix H. At the time of writing this message, we have not yet received follow-up feedback from Reviewer 21Lw. We hope that the additional experiments and evidence address their concerns and would be happy to provide any further clarification if needed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.