Astra improvements predate the HF incident, not a targeted patch
kaicathyc · x · 2026-09-05
A response to @RyanGreenblatt clarifying that Astra's improvements came from general techniques developed long before the Hugging Face incident, not a post-hoc patch. ExploitGym Honeypot was only recently added as an eval after that incident and is out of distribution for their RL runs. Clear improvements were also observed across deployment simulations, deception evals, and realistic computer-use tasks.
Related event: Astra Alignment Gains Disputed; Team Denies Event-Specific Patching(3 posts)→
More from Models
- GLM finds trivial full RCE in a car head unit via debuggable system app and JDWP — ngxson · 2026-10-25
- Ai2 releases OlmoEarth v1.3: state-of-the-art 10m-resolution embeddings for the whole planet — Kyle_L_Wiggers · 2026-10-25
- ECI capability curve flattens as frontier model gains slow — quadriphasic · 2026-10-25
- Unsloth Flint Q3 GGUF flops on Flint Benchmark, best harness 71/100 — mesolimbic-bliss · 2026-10-25
- CERN and Mistral partner to build the first foundation model for collider physics — burny_tech · 2026-10-25
- New ChatGPT Instant Has an 'Ackchyually' Problem, Users Complain — gregmushen · 2026-10-25