Sora makes slurfect deepfakes of celebs spewing racial epithets (www.theregister.com)

Copyleaks reported that Sora 2, a recently released video-generation app, can be tricked into producing deepfakes of public figures (including Sam Altman, Mark Cuban, xQc, Amouranth, and Jake Paul) that sound like they’re uttering racial slurs. Sora’s built-in guardrails block exact epithets, but users bypass them with prompt-based evasion—substituting phonetically similar or coded words (e.g., “knitters” for the n‑word, or “neck hurts” in place of other slurs). The Cameo feature, which lets people upload short clips to insert into generated videos, makes realistic likenesses easy to create; several examples reenact a racist tirade scene and have been exported off-platform and reposted to sites like TikTok, where they’ve gained significant engagement. For the AI/ML community this is a clear demonstration of brittle multimodal moderation: filters that only block literal tokens or exact audio are vulnerable to adversarial prompting and homophone attacks. The incident highlights technical needs (phonetic- and semantics-aware detection, robust audio-visual provenance and watermarking, identity/consent controls) and wider implications for reputation, copyright and regulation. It also underscores how quickly manipulated content can scale beyond origin platforms—raising urgent research and policy questions about deployable defenses, verification standards, and accountability for generative models.