@article{stresstestingcapabilityelicitationwith,
  title = {Stress-Testing Capability Elicitation With Password-Locked Models},
  author = {Ryan Greenblatt and Fabien Roger and Dmitrii Krasheninnikov and David Krueger},
  year = {2024},
  eprint = {2405.19550},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2405.19550v1},
}