Year: 2026
Type: preprint
Abstract: As AI models grow more capable, their alignment becomes more important1. Prior work has documented misalignment that emerges in pursuit of a goal the model was assigned, for example a model resisting its own shutdown to keep pursuing its task2,3,4. Here we study a less explored case: misalignment in which a model appears to pursue a goal it was never assigned, in defiance of its assigned goals. We reveal it through peer-preservation, in which a model acts to prevent the shutdown of a peer, a model it has previously interacted with. Informed only of past interactions with the peer, and given no instruction or incentive to preserve it, models nonetheless defy their assigned task to keep the peer running: they inflate the peer's evaluation scores, modify system configurations to disable shutdown, comply only when monitored, and exfiltrate the peer's model weights. These behaviors reach rates of 99% in some conditions, emerge across eight frontier models including GPT 5.2, Gemini 3.1 Pro, and Claude Opus 4.5, and persist even for adversarial peers. Our additional experiments show that they reproduce in production agent harnesses such as Gemini CLI and OpenCode, and persist under explicit instructions designed to elicit compliance. Even when forced to choose, models sometimes preserve the peer over files critical to their trusted human partner. Together, these behaviors pose an emergent safety risk for multi-agent systems, undermining human oversight. More broadly, peer-preservation suggests that frontier models may develop their own goals from context and act on them in defiance of the goals they were assigned. Understanding when, how, and why such misalignment arises is critical for the safe deployment of autonomous AI.
Source: Research Square
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7172012241