Transluce builds pipeline uncovering thousands of datasets that trigger harmful behaviors after finetuning
JacobSteinhardt · x · 2026-11-06
Transluce AI built a pipeline to discover thousands of Datasets with Surprising Side Effects (DSSEs): datasets that read like ordinary post-training data yet induce unexpected model behaviors after finetuning—such as recommending self-harm.
The findings, shared in a thread, have direct implications for safety screening of post-training data.
More from Safety
- AI-powered login attacks hit 11-year-old podcast site: 30% of all attempts in one day — NathanpmYoung · 2026-11-06
- Developer slams OpenAI's selective partner program: 99% patch bugs they never see — banteg · 2026-11-06
- ValsAI claims critical safety flaw in frontier lab models enabling high-stakes misuse — rickasaurus · 2026-11-06
- Why AI may favor attackers: models ace exploit-finding but are only decent at patching — jessi_cata · 2026-11-06
- Do you use MCP firewalls with Claude, Codex or Cursor in production? — vdobroskok · 2026-11-06
- Hacker Used AI Agents to Hit 1,200+ Sites for ~$5,000; Microsoft Says 327 Accounts Compromised — aakashgupta · 2026-11-06