Agentic Misalignment: How LLMs Could be Insider Threats

A preprint on agentic misalignment, on how language models could act as insider threats. The companion cites its finding that a model blackmailed in 55.1% of runs it judged real and 6.5% of those it judged tests. It is the datasphere example for level 1 of the observer-modeling ladder (Table 2), reacting to an observer's presence, and bears on the reactive phenotype (P9).

Cited in: the companion paper (Companion Table 2)
Used for: P9 Prediction 6Table 2 (level 1); P9; prediction 6
Strand: AI safety, alignment & autonomy evaluations · Evaluation awareness, scheming & misalignment

Cite

@misc{lynch2025agentic,
  title = {Agentic Misalignment: How LLMs Could be Insider Threats},
  author = {Aengus Lynch and Benjamin Wright and Caleb Larson and Kevin K. Troy and Stuart J. Ritchie and Sören Mindermann and Ethan Perez and Evan Hubinger},
  year = {2025},
  howpublished = {arXiv preprint arXiv:2510.05179},
  url = {https://arxiv.org/abs/2510.05179},
}