Definition
Plain language
Having full access to a system's insides, not just its outputs.
As stated in the literature
An analysis or attack setting with access to weights and internal activations; a white-box linear probe on hidden states can score per-answer fatality at high AUROC even when the outputs themselves look uniformly compliant.
Also called: white box, white-box probe
Why it matters: It marks the difference between studying a system from the outside and being able to see what it is actually doing internally, which changes what both defenders and attackers can achieve.
For example, a researcher with the full model can watch its internal signals while it answers and tell which responses are secretly wrong, even though every answer looks equally helpful on the surface.
Heard on the show
“That's a reasonable scientific compromise — you can't open up GPT-4o — but it's worth being honest that the white-box claim is narrower than the black-box claim.”Episode 073 — When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving