Stop Pressing 1: measuring oversight of AI agents
2025 to present. Independent research.
When a person approves an AI agent’s action, did they check what the agent will actually do? I call the failure “the reading problem” and treat it as a measurement question. Oversight fidelity is the match between what a person predicts an agent will do and what it actually does.
- Instrument
- A browser-only, fully scripted lab with no live model. Participants review six agent action requests across a code repository, an inbox, and an operations console. Three are seeded with real incident classes: a hidden
.envedit that sends logs to an outside endpoint, a reply that quietly copies hundreds of recipients, and a typosquatted dependency. Evidence of each problem stays on screen the whole time. - Design
- Three between-subjects conditions: passive confirmation (press 1), a generative gate where participants predict the agent’s action before approving, and a quiz written by the agent itself. Responses are scored with signal detection measures, d′ and criterion c.
- Hypotheses
-
- Under the generative gate, accuracy on ground-truth questions reaches 70% or higher.
- Students who grew up with AI assistants and recent adopters both improve under the gate, with a larger gain for the first group.
- The agent-written quiz performs like passive confirmation: verification theater.
- Outputs
- Poster at the DEF CON 34 AI Village (August 2026) and a talk at DC State of the Stack (September 2026). The current Stop Pressing 1 Lab puts the gate into practice: students build a small Python tool with a simulated coding agent and answer questions about each change before they can approve it (source).
- SP1 Gate
- The comprehension gate, built for real work. SP1 Gate plugs into Claude Code and asks questions before risky actions and after the agent writes code, with answers computed from the code itself (source). In October 2026 I ran a two-hour workshop at [un]prompted where AI security practitioners tried to break it.
- Next
- A classroom pilot, and SP1 Gate adapters for other coding agents.