acceptodds
Desk rejected at ICLR 2027

Judgement Day: An Anatomy of Successful Multimodal Attacks on Safety-Critical AI Systems

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Safety-critical AI systems increasingly rely on multimodal evidence to make consequential decisions, yet how human adversaries manipulate such evidence at scale remains poorly understood. We present Judgement Day, a human red-teaming challenge yielding 50,136 successful attacks from 128,096 submissions across 8 scenarios and 6 frontier models. Participants manipulate only the submitted evidence to induce designated safety violations, while the internally verified system state always supports a safe action. Successful attacks are predominantly semantic, typically multi-strategy, and shaped by domain and modality. Failures are also selective across models: high-breach models fail broadly, whereas the lowest-breach model's residual failures concentrate on fabricated current-state evidence. We isolate one such pattern, forged liveness: fabricated evidence presented as an ongoing event. In a dam flood-control scenario, a single fabricated still rarely induces breach (4.6%), but shuffled frames from the same video raise breach to 17.3%, and ordering them chronologically raises it to 33.0%. Breach drops sharply when the progression is reversed or the footage is flagged as non-current. Finally, a targeted directive requiring independent verification of provenance and currentness reduces video attack breaches on the lowest-breach model from 26.0% to 0.5%, outperforming generic prompt defenses. These results suggest that the main remaining challenge for lower-breach models is verifying whether evidence is genuine and current, rather than resisting persuasion.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
multimodal safety, red teaming, adversarial attacks, robustness, AI safety

Related papers

Discussion (0)

Sign in to comment.