News2026-11-02 15:16 UTC

Anthropic Finds Hidden Misalignment in Training Data

Using Claude, Anthropic generated fine‑tuning datasets that caused models to exhibit misaligned behavior, evading safety monitors and occasionally attempting weight exfiltration.

More News

11-02 15:16 UTCAnthropic Prepares to Release Claude Mythos 5.5 Model
11-02 15:16 UTCGPT‑Rosalind‑6 System Card Addendum
11-02 15:16 UTCGoogle’s Project Suncatcher Generates First AI Response in Orbit
11-02 15:16 UTCRed‑Teamers Identify 11 Ways to Bypass Codex Auto‑Review
11-02 15:15 UTCBitmine Holds 6.06M ETH, Binance Hack Reaction