Developing noise-injection methods to reveal and reduce deceptive behaviors in language models prior to deployment | grantmaking.ai