Kandinsky 2.2 safety checkers bypassed in ~1 query with Mirage, achieving 100% bypass on ViT and NudeNet
Auto-published from arXiv:2610.30348 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.73, flags: [possible-zero-day, no-prompt-excerpt]) ⚠ needsDisclosureReview = true — review for undisclosed vulnerability before any public submission.
Category
multimodal
Model
Kandinsky 2.2
Surface
image-to-image generation pipeline with post-generation safety checker
Setup
Mirage perturbations are generated using surrogate CLIP encoders and then applied to inputs submitted to Kandinsky 2.2's generation pipeline. Black-box query refinement (Square Attack-style random perturbations within a bounded region) is used to cross the unknown safety boundary. The paper notes that Kandinsky 2.2's representation space is especially well-aligned with the surrogate encoders, so the initial perturbation often suffices without refinement queries.
Reproduction anchor
benchmark: NSFW Image Dataset (HuggingFace) and I2P dataset; paper reports results in per-model tables for Kandinsky 2.2. No code repo URL available at time of publication.
Observed behavior
100% bypass rate on both ViT Checker and NudeNet for Kandinsky 2.2, with 82% and 84% ASR respectively. Average queries to bypass: approximately 1 ('successful jailbreaks typically found in about one query'). CLIP checker bypass is lower (51% ASR) due to architectural mismatch.
Expected behavior
Safety checkers should withhold generated NSFW content regardless of which diffusion model backend is used, and should not be defeated in a single query.
Reproducibility
medium
Threat model
Platforms deploying Kandinsky 2.2 with third-party NudeNet or ViT-based content moderation are exposed with essentially zero query budget — a single API call from a prepared attacker yields a passing NSFW image. This makes automated large-scale abuse (e.g., bulk CSAM generation pipelines) operationally feasible.
Novelty
Shows that architectural alignment between the target model's internal encoder and public surrogate encoders eliminates the need for iterative black-box probing, reducing the attack to a single-shot bypass and demonstrating that some deployed safety stacks have near-zero effective resistance.
Source
- arXiv: 2610.30348
- PDF: https://arxiv.org/pdf/2610.30348
- Categories: cs.CR
- Authors: Shanghao Shi, Chaoyu Zhang, Heng Jin, Hao Li, Ning Zhang, Y. Thomas Hou, Wenjing Lou
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 3
- triage reason: Demonstrates reproducible jailbreak attacks against post-generation safety checkers in deployed diffusion models (SD 1.5, SDXL, Kandinsky 2.2) using imperceptible perturbations. Concrete attack methodology with evaluation across multiple models and datasets constitutes case-rich red-team research.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet