Banger paper from Meta FAIR on catching distilled models.
This is a dev post classified by AGK Intelligence as Security (a free resource), kept by the Dev Radar because it carries real work, not commentary.
Banger paper from Meta FAIR on catching distilled models. (bookmark it) At a 0.1% false-positive rate, their detector flags 79% of models distilled from watermarked outputs. The best token-level test flags 29%. When a student trains on watermarked teacher traces, the watermark key leaves a specific direction in the student's weights. That direction sits apart from the normal fine-tuning directions, so it stays even when the student's text shows no token-level trace of the watermark. WM-Basin only needs API access. It turns the suspect model's outputs into gradients on a clean probe model,
Posted by DAIR.AI (136.1k followers) 1 days ago · 118 likes · 7.6k views · view the original post on X. Kept by the Dev Radar as Security.
More dev work like this
- CVE-2024-21413: MonikerLink Outlook Vulnerability PoC — @DarkWebInformer
- CrowdStrike President Michael Sentonas joined Sascha Roehrer on deeptech INSIGHTS to… — @CrowdStrike
- I just shared this with my family, and you all should do it too. — @JJEnglert
- Securing AI shouldn't add complexity. — @PaloAltoNtwks
- Standing room only at the Boston Kubernetes Meetup last night for Teleport’s Paul Curtis… — @goteleport
- AI adoption is accelerating. Stopping emerging AI threats takes elite expertise. — @CrowdStrike
- Names, emails, API keys—Our ambassador @povedaai is putting MiniCPM4-0.5B to work… — @OpenBMB
- Finding vulnerabilities is one thing. Preventing them from being merged is another. 🔐 — @gitlab
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.2k posts from 5k X accounts over the last 21 days, 8.8k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-10-30 20:27 UTC. Full method.