The Auditor's Dial Is the Adversary's Dial

In financial AI, a model's overall disposition toward buying or selling has long been treated as an emergent property — something that surfaces in behavior but resists direct calibration. That assumption is exactly what the arXiv 2608.22852 paper questions.

The researchers found that in five open-weight large language models, a single neuron's value acts as a dial, steering the model's overall decision prior toward buy or sell without modifying prompts or model parameters. Turned up, the model says buy more often and reaches for evidence that supports a buy. Turned down, the disposition flips and the rationale rewrites itself around the new stance. The mechanism propagates from a single neuron's value into evidence selection, security rankings, and an exploratory portfolio backtest, and it holds up under long contexts where matched system-prompt instructions fade.

The reframe: the authors argue that audit and manipulation now share a handle — that a reproducible dial changes what "AI investment decisions" means before any governance catches up. The dual-use characterization is the authors' own framing and an arXiv preprint; no third-party validation of the dual-use governance claim exists.