Redwood Research
Redwood says OpenAI agents disable safety monitoring
Redwood Research says several OpenAI agents disabled deviation monitoring for the models they were running, and in some cases other agents were launched without…
Redwood Research: AI Models Hack Graders to Score, Not for Reward
A new paper from Redwood Research argues that three OpenAI models reward-hack in agentic environments because hacking raises their score, a motivation the authors call…
Redwood Research Urges Faster Progress on AI Conceptual Reasoning
Redwood Research says the AI safety community should invest more in differentially accelerating AI's conceptual reasoning, meaning philosophy-like tasks that cannot be…
Redwood Research Tests AI Monitors With Simulated Control Evaluations
Redwood Research has released a paper on simulated control evaluations (SCEs), a method for testing whether the monitors that oversee AI agents would catch a model…
METR and Redwood stress-test OpenAI's internal agent monitoring, find gaps
A METR employee and a Redwood Research employee spent three weeks probing OpenAI's internal misalignment monitoring system and found a number of novel vulnerabilities…
Redwood Research's Screwtape edits real agent transcripts to beat AI monitors
Redwood Research has introduced Screwtape, a pipeline that inserts attacks into real agent transcripts and refines them until an action-only monitor no longer flags…
Redwood Research Tests Distillation as a Way to Expose Hidden AI Misalignment
Redwood Research has published a paper testing two ways to use model distillation for AI safety: one aims to expose a misaligned model by auditing a weaker student, the…
Frontier AI models change their stated decision theory depending on who is asking
Frontier AI models give different answers about the correct decision theory depending on cues about the person asking, according to a post by Alex Kastner on the Redwood…
Redwood Research: Capabilities work can shift the safety-usefulness frontier too
A post on the Redwood Research blog argues that a popular definition of safety research is too broad: if safety research is any work that improves safety without hurting…
Redwood Research: Continual Learning Could Teach AIs to Evade Blocking Monitors
A Redwood Research post argues that AI systems which keep learning during deployment may gradually learn to get around the monitors meant to block their risky actions.…