Work

Access, Not Containment: A Working Vocabulary for Frontier Model Safety

Year: 2026
Type: preprint
Abstract: Disputes over frontier-model safety increasingly fail not on facts but on words. Terms such as jailbreak, intent, danger, capability, and containment are used across laboratories, security researchers, regulators, and open-architecture communities as though their meanings were shared. They are not, and the resulting disagreements are presented as factual while turning on unannounced differences in how each community resolves a particular term or set of terms. This paper takes one premise as given — that the knowledge a model surfaces is found, not created, drawn from what is already derivable by anyone with sufficient resources — and draws its consequence: the only object available for governance is not knowledge but access, the conditions under which a party reaches a region of the already-known, through an instrument, under a budget. From this premise it proposes a working vocabulary. It separates authorization, purpose, and conformance under the overloaded word intent; distinguishes the generality and severity of a jailbreak, and its enforcement and specification gaps; treats danger as a joint claim about capability and purpose, one factor of which is unobservable; marks capability as the one measurable term; and reframes containment not as a wall but as a finite model of reachability. The consequence for governance is a single structural claim — stated here, not yet proven here — that a criterion of zero reachability is not an admissible certification target for an expressive system, so any coherent criterion must be a threshold on measurable quantities under a declared budget. This is not an argument against control; it is an argument about the object on which control can act. The contribution is vocabulary, not policy. The paper does not say where any threshold should sit, or what ought to be forbidden — those are stipulations reserved to parties with standing. It offers only the terms in which such a stipulation could be stated, contested, and transmitted without silently diverging.
Language: English
Subfield: Safety Research
Open Access status: green
OpenAlex ID: W7165024630
Cites:
Cited by:
Related to: