keshavs

Karma: 55

Attackers Can Subliminally Implant a Backdoor at Low Sample Count Without Prompt Access

keshavs3 Aug 2026 22:08 UTC

23 points

3 comments5 min readLW link

Introspection Adapters: Training LLMs to Report Their Learned Behaviors

keshavs, RowanWang, abhayesian and Sam Marks

28 Apr 2026 19:02 UTC

41 points

2 comments12 min readLW link

(alignment.anthropic.com)