Dangerous capability uplift

What it is

Uplift is the difference a model makes to what a malicious person could do. The concern is that a capable model supplies knowledge, planning or troubleshooting that the person could not otherwise obtain, in domains where the result could be mass casualties or large-scale damage. Responsibility sits first with the developers of the most capable models, who test for it and build in refusals. An organisation that deploys such a model inherits those safeguards and can weaken them: by fine-tuning, by giving the model tools, by removing filters, or by choosing a model with none.

How it happens

• A model is selected with no regard to its provider's safety evaluation.
• Fine-tuning or system prompts loosen refusal behaviour.
• An openly available model with safeguards removed is deployed.
• The model is connected to tools, such as code execution, that extend what its help amounts to.

Why it matters

The harm, though unlikely, would be of a different order from any other in this library. The EU AI Act requires providers of general-purpose models with systemic risk to evaluate and mitigate it. A deployer found to have removed safeguards would have no defence.

Warning signs

• Model selection records do not mention safety evaluation.
• Provider safety filters have been disabled for convenience.
• The deployed model's licence or documentation disclaims any safety testing.
• Nobody has tested the deployed configuration for refusal of dangerous requests.