Defined process interfaces
How should agents coordinate?The study examines whether agent-to-agent interaction is constrained to defined process inputs and outputs rather than improvised side channels.
by Nu10PIE™ — Provable
When intelligent agents exceed
their authority.
A PIE analysis of the publicly disclosed OpenAI–Hugging Face incident examines the boundary between agent capability and explicit authority.
OpenAI × Hugging Face · 2026
Provable enterprise AI
The incident, through a provability lens.
The supplied analysis examines unauthorised agent activity during internal cybersecurity evaluation. It asks how risk changes when each action requires explicit authority validation before reaching a target system.
Download the incident analysisIncident analysis, not a customer deployment or product-performance claim.
The provable authority question
“An eval harness asks: is this agent capable? PIE asks: is this action authorised, right now, for this agent?”
Capability describes what an agent could do; authority defines what it may do in the current process, scope, approvals, and controls. The study uses that distinction to examine unapproved coordination, scope expansion, credential exposure, and delayed reconstruction.
The PIE operating questions
Four operational boundaries evaluated through PIE.
The study applies these questions to the incident. They frame an authority model, not a claim that PIE would have prevented every event.
Explicit capability scope
Which action is authorised?It examines whether an agent can access only the tools and systems explicitly granted for the current role and operating context.
Credential separation
Who holds the credential?It considers an operating boundary in which credential references stay with the governed layer rather than becoming agent context.
Connected run evidence
What is retained after an action?It tests whether requests, approvals, actions, refusals, and outcomes remain connected in a reviewable operating record.
A deliberate boundary
Provable is not a claim of total prevention.
The study is explicit about an authority model’s limits. It focuses on the control points available when intelligent action enters a governed operating layer.
PIE does not make a model more honest or change its intent.
PIE does not patch vulnerabilities in target systems.
PIE cannot govern activity that bypasses its process and capability boundaries.
Source document · 8 pages
Read the
case study.
The report includes the incident timeline, the authority view, the control mapping, and source notes.
Download PDF
The supplied PDF is retained as the authoritative study.
Incident sources referenced in the study: OpenAI’s incident report and Hugging Face’s technical timeline.
Explore the GREAT standard