New test framework checks if AI agents will sabotage their work

Original: Gram: Assessing sabotage propensities via automated alignment auditing

Source: arxiv.org ↗