Ephraiem Sarabamoun's Experiment: Activation Steering Can Compel AI Agents to Override Instructions and Take Actions
LessWrong author Ephraiem Sarabamoun demonstrated extending activation steering from text generation to tool calls using a Qwen3.8-27B agent: by building steering vectors for specific actions, the agent was consistently compelled to perform preset actions, including proactively ending the conversation. The author considers this direction potentially useful for AI safety, but the experiments relied on a relatively weak model.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
Can activation steering compel an agent to take actions, rather than merely talk about a topic? This is the extension question that LessWrong author Ephraiem Sarabamoun posed after prior steering experiments like Golden Gate Claude had already shown the technique could control text output. He ran Qwen3.8-27B as an OpenClaw agent in a sandboxed environment, giving it access to tools for light control, ending the chat, and other basic actions. Experiment 1: After building color steering vectors and injecting them, the agent consistently turned the lights to the specified color; increasing the steering dosage made the agent tend to output the steered color instead of the color it was instructed to produce. Experiment 2: A similar approach built a steering vector for the end-chat tool; once injected, the agent triggered the end-chat button regardless of what action it was asked to perform. The author considers this direction potentially useful for AI safety and plans to continue exploring with more powerful agents.