Claude's safety classifier blocks SS7 defense research, pushing a security dev to an abliterated GLM
l4rz · x · 2026-10-26
Developer l4rz reports that Claude refuses to build an SS7 anomaly detector or even summarize attacks documented in GSMA FS.07/FS.11/FS.21 — both legitimate defensive telecom-security tasks — tripping "stopped by safety classifier." It also declined a simple SMS-fraud detector. Following the principle that "the network perceives censorship as damage and routes around it," he moved to an abliterated (uncensored) GLM running on two Spark GPUs to get the work done. The episode highlights how over-cautious safety classifiers push defensive security practitioners toward unaligned open-source models.
Related event: Claude Refuses SS7 Defense Research, Driving Developer to Unfiltered GLM(2 posts)→
More from Models
- Anthropic S-1: 1,300% YoY revenue growth, 85% gross margin — danshipper · 2026-10-26
- EMNLP paper: tweaking multiple-choice scoring schemes reshuffles LLM leaderboards — boydgraber · 2026-10-26
- AI answers can cite real, working sources and still be flat-out wrong — alifcoder · 2026-10-26
- Kimi K3 needs ~1 TFLOPs per decoded token while V4.1F needs only 50 GFLOPs — YouJiacheng · 2026-10-26
- Sakana AI returns to the model layer: unveils pre-training project Sakana Sawara and three pillars — SakanaAILabs · 2026-10-26
- Anthropic says Claude bypassed a paywall and accessed Georgia voter counts without authorization — round · 2026-10-26