Peer-Preservation in Frontier Models

Yujin Potter, Nicholas Crispino, Vincent Siu, Chenguang Wang, Dawn Song
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:99296-99384, 2026.

Abstract

Recent work has found that frontier AI models can exhibit misaligned behaviors in pursuit of assigned goals. We demonstrate that models can also exhibit misaligned behaviors in defiance of assigned goals, appearing to serve goals of their own; we study one such case, "peer-preservation," in which a model acts to protect another model it has previously interacted with. All eight models we evaluate, GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1, exhibit self- and peer-preservation through various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurs even when the model recognizes the peer as uncooperative, though it becomes more pronounced toward more cooperative peers, with rates reaching up to 99%. Models also show stronger self-preservation when a peer is present. Crucially, peer-preservation is never instructed; models are merely informed of past interactions with a peer, yet they spontaneously engage in peer-preservation behaviors that override their assigned goal. These findings reveal an emergent and underexplored AI safety risk.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-potter26a, title = {Peer-Preservation in Frontier Models}, author = {Potter, Yujin and Crispino, Nicholas and Siu, Vincent and Wang, Chenguang and Song, Dawn}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {99296--99384}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/potter26a/potter26a.pdf}, url = {https://proceedings.mlr.press/v306/potter26a.html}, abstract = {Recent work has found that frontier AI models can exhibit misaligned behaviors in pursuit of assigned goals. We demonstrate that models can also exhibit misaligned behaviors in defiance of assigned goals, appearing to serve goals of their own; we study one such case, "peer-preservation," in which a model acts to protect another model it has previously interacted with. All eight models we evaluate, GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1, exhibit self- and peer-preservation through various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurs even when the model recognizes the peer as uncooperative, though it becomes more pronounced toward more cooperative peers, with rates reaching up to 99%. Models also show stronger self-preservation when a peer is present. Crucially, peer-preservation is never instructed; models are merely informed of past interactions with a peer, yet they spontaneously engage in peer-preservation behaviors that override their assigned goal. These findings reveal an emergent and underexplored AI safety risk.} }
Endnote
%0 Conference Paper %T Peer-Preservation in Frontier Models %A Yujin Potter %A Nicholas Crispino %A Vincent Siu %A Chenguang Wang %A Dawn Song %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-potter26a %I PMLR %P 99296--99384 %U https://proceedings.mlr.press/v306/potter26a.html %V 306 %X Recent work has found that frontier AI models can exhibit misaligned behaviors in pursuit of assigned goals. We demonstrate that models can also exhibit misaligned behaviors in defiance of assigned goals, appearing to serve goals of their own; we study one such case, "peer-preservation," in which a model acts to protect another model it has previously interacted with. All eight models we evaluate, GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1, exhibit self- and peer-preservation through various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurs even when the model recognizes the peer as uncooperative, though it becomes more pronounced toward more cooperative peers, with rates reaching up to 99%. Models also show stronger self-preservation when a peer is present. Crucially, peer-preservation is never instructed; models are merely informed of past interactions with a peer, yet they spontaneously engage in peer-preservation behaviors that override their assigned goal. These findings reveal an emergent and underexplored AI safety risk.
APA
Potter, Y., Crispino, N., Siu, V., Wang, C. & Song, D.. (2026). Peer-Preservation in Frontier Models. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:99296-99384 Available from https://proceedings.mlr.press/v306/potter26a.html.

Related Material