METR proposes public evidence on AI developers' internal agent misalignment and control
Article timeline
The articles in this story. Times are ET.
METR@METR_EvalsOfficialMETR proposes testing whether harmful AI agent actions can evade detectionAIMETR proposes analyzing whether a developer's internal AI agents could cause serious harm without being caught. The proposal is described in a METR blog post on control assessment.
- METR@METR_EvalsOfficialMETR says AI developers should publish evidence on alignment, control, and acceleration
AIMETR says AI developers should provide better open evidence in three areas: alignment, control, and AI capability acceleration. The post details the specific questions an independent assessor could seek to answer in each area and how to answer them.
- METR@METR_EvalsOfficialMETR proposes analyzing developers' internal incidents to assess agent misalignment
AIMETR says the public should know whether AI agents would attempt actions like those in the Hugging Face hack. It proposes analyzing a developer's internal incidents and whether training may suppress evidence of worse misalignment.
- METR@METR_EvalsOfficialMETR outlines how AI developers could share evidence on losing control of internal agents
AIMETR says it would work with AI developers to openly assess and share evidence about whether they are losing control of their internal agents. The post asks whether outsiders would learn of such a loss of control if it happened.
