Evidence File — Who Gets Woken Up
Evidence file · published with the report
Primary documents
| Document | Role | Use |
|---|---|---|
| https://openai.com/index/how-we-detect-frontier-model-misalignment/ | Primary source | Supports the article's claims |
| https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/ | Corroborating source | Institutional page, same publisher |
Claim-by-claim verification
| Claim as published | Status | Basis |
|---|---|---|
| OpenAI published the monitoring description on 29 October 2026 | CONFIRMED | Page byline, "OpenAI October 29, 2026" |
| The system has four automated layers plus a human response stage | CONFIRMED verbatim | Post's contents list: Layer 0 activation probes through Layer 3 automated investigation, then Human response |
| Paging is currently disabled for internal Codex traffic | CONFIRMED verbatim | Load-bearing quotation, verified twice against the retrieved page |
| External monitoring does not routinely escalate flagged cases to human reviewers at OpenAI | CONFIRMED verbatim | "Unlike our internal system, our external misalignment monitoring does not routinely escalate the cases it flags to human reviewers at OpenAI." |
| External monitoring does not cover Chat Completions API requests | CONFIRMED verbatim | Confirmed on a second full retrieval, 2 November 2026 |
| Without conversation-tracking or webhooks, an alert can arrive and the conversation continues | CONFIRMED verbatim | "Without those mechanisms, a configured webhook can still receive an alert, but the conversation can continue." |
| A broader 30-minute paging safeguard does not yet exist | CONFIRMED verbatim | "We're working toward a broader safeguard that would pause affected activity if responders have not ruled out a highest-priority alert within 30 minutes of being paged." |
| The 30-hour figure is a simulation, not an observed alert | CONFIRMED | "In that simulation, our monitoring system alerted our security team" |
| The original coverage gap concerned unmonitored internal cybersecurity evaluations | CONFIRMED verbatim | "began during internal cybersecurity evaluations on which we were not running our monitors for misaligned behavior" |
| Frontier RL runs and internal Codex traffic skip the 5 percent sampling layer | CONFIRMED verbatim | "we have our frontier RL runs and internal Codex traffic skip this layer and proceed directly to the next one" |
| METR and Redwood red-teamed the stack for three weeks from July 2026 | CONFIRMED verbatim | "METR and Redwood Research spent three weeks, starting in July 2026" |
| Most red-team findings are patched; one fix is still in progress | CONFIRMED verbatim | "Most of these have now been patched; one remaining fix is still in progress." |
| OpenAI recommends all developers monitor their most capable models | CONFIRMED verbatim | "We believe that all model developers should monitor the behavior of their most capable models." |
Known limits
- Single primary source. The article describes OpenAI's system from OpenAI's own document; no independent observation of paging, escalation or webhook behaviour exists.
- No interview. OpenAI was not contacted for comment, and the article says so.
- The article does not claim the monitoring system is ineffective. It argues the opposite in terms, and names the compensating control.
- No claim of exclusivity, and no claim about how widely the story was covered.
- The 30-hour figure is presented as the simulation OpenAI says it is; the article does not treat it as an observed alert.
Provenance
Candidate identified by the researcher agent; every quotation independently re-verified by COMPUTER against the retrieved page text before drafting. Revised on 2 November 2026 after an independent review that retrieved the primary source itself (7 corrections applied, including a logic inversion about the external action path and the addition of the Chat Completions carve-out, which was confirmed verbatim on a second retrieval). Converted to this package by build-package-generic.ps1 rather than assembled by hand. Staged through stage-package.ps1. published, not committed, not deployed.
Produced under human editorial authority. Nothing in this file is published.