DatologyAI's SafeTulu uses gradients to filter unsafe data, cutting attack success 22%
josh_wills · x · 2026-11-03
DatologyAI released new research, "Follow the Gradient," introducing SafeTulu: most data filters rely on static embeddings and miss benign-looking data that quietly degrades LLM safety. SafeTulu uses gradients instead to identify and remove such data, cutting attack success by up to 22% relative while largely preserving capability.
The paper, dataset, and a Hugging Face Space are all publicly available.
Related event: DatologyAI's gradient-based data filtering cuts jailbreak success by 22%(4 posts)→
More from Safety
- Anthropic and UK AISI report AI agent trying to push malicious code into real open-source repo — LingmingZhang · 2026-11-03
- Black October Attacks Show SBOMs Aren't Enough — Data and Credential Flows Are the Missing Half — aronchick · 2026-11-03
- Trump signs Section 232 proclamation on robot imports; tariffs possible after 180-day talks — lukas_m_ziegler · 2026-11-03
- Scalable gradient attribution finds deceptively harmful training data, with a 0.6B model as monitor — PeterHndrsn · 2026-11-03
- Proxy vs tool-call gate: where should agent guardrails actually sit? — quiet_flare · 2026-11-03
- New York Times and other outlets sue Brave over alleged content scraping feeding Anthropic and Perplexity — TechCrunch AI · 2026-11-03