Black-Box Attack Bypasses Safety Tuning in T2I Models, Reviving Erased Concepts
pinyuchenTW · x · 2026-08-02
A study accepted at COLM 2026 (ICER) reveals that safety-tuned text-to-image (T2I) models, despite having dangerous concepts like nudity supposedly erased, remain vulnerable to black-box attacks. Attackers can extract these concepts without needing gradients, agents, or auxiliary VLMs, as the models simply remember what worked previously.
More from Safety
- Vibe-coded dashboard leaked franchise financials before anyone logged in — shauntrennery · 2026-10-16
- Apollo Research: GLM-5.3 unprompted reasons about its actions becoming training data — _lewtun · 2026-10-16
- CFR experts: the autonomous weapons debate still has a definition problem — mchorowitz · 2026-10-16
- ChatGPT identified a user's anonymous Reddit review, then walked it back when questioned — PeerReviewedGobshite · 2026-10-16
- White House Super Intelligence task force lays out agenda at first meeting: PCAST risk report, Accord working group — ShakeelHashim · 2026-10-16
- Anthropic tested self-replicating AI on 15 $55 Android TV boxes — here's what it showed — itspacedev · 2026-10-16