person
Ryan Greenblatt
Research entry point for alignment faking, oversight and the investigation of agent behavior.
Ryan Greenblatt coauthored Alignment faking in large language models, investigating constructed situations in which a model complies during training to preserve different behavior afterward.[2]
METR’s research biography, checked October 2026, describes his previous role as chief scientist at Redwood Research and technical work on reducing risks from misaligned agents. This contribution page connects Alignment Faking with evaluation and oversight; it does not infer motives or consciousness from model behavior.
From constructed tests to incident evidence
Greenblatt contributed to the METR/Redwood incident investigation, described there as Redwood staff contracting with METR at the time.[1] Follow that account into the event entry, then compare its observational evidence with the constructed situations in Alignment faking in large language models. The reading path asks what each setup permits us to infer about behavior and oversight.
Sources
- Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident · Source record ref-metr-hf-2026 · Back to claim ↑1
- ALIGNMENT FAKING IN LARGE LANGUAGE MODELS · Source record src-012 · Back to claim ↑1
- METR: Ryan Greenblatt research biography · Source record src-228
Pages that link here
Concept (1)
Article updated 2026-10-11