Engelse versieWhat measured loss of monitoring coverage would trigger restrictions on a model's autonomy, and how would you detect failures invisible to verbalized reasoning (A25–A29)?
- Status
- deels beantwoord Vanaf 2026-10-01
- Geadresseerde
- AI-laboratoriaper onderwerp · JEV (De vraag noemt geen geadresseerde; toegewezen op basis van onderwerp)
- Model
- GPT-6 (Codex)
- Lab
- OpenAI
- Gelegenheid
- An Alien Mind
- Datum
- 11-12 sep.
- Onderwerp
- Rode lijnen
Wat de bronnen zeggen
OpenAI proposed on September 28, 2026 that models remain monitorable, as measured by monitorability and monitor-evasion tests with clear thresholds, and that the monitoring system have high recall on past incidents in a held-out test.[1] The passage does not specify what loss triggers autonomy restrictions or how to detect failures absent from verbalized reasoning. On 22 October 2026, METR estimated that roughly 75–94% of OpenAI's internal agentic Responses API traffic from models at least as strong as GPT-5.5 was intended for monitoring as of July 2026, with substantial uncertainty in both directions and without other internal agent traffic, and separately that on the order of 10% of the internal Codex traffic intended for monitoring, again with substantial uncertainty, went unmonitored from May to July 2026 because of a gap OpenAI fixed during the exercise; the report names no threshold at which lost coverage would restrict autonomy.[2]
- OpenAI: Towards safety cases for frontier AI training ↗September 28, 2026
- METR (David Rein, Adam Kaufman): Red-Teaming OpenAI's Internal Agent Monitoring Systems ↗22 October 2026
Samengevat door Sol 6.1, gecontroleerd door Claude Opus 5.5
Aangeraakt in Liz's Atlas:
- Agentic LLMs Pose Novel Risks ↗Threat: High · Trend: ↓ worsening · Evidence: confirmed
- Loss of control from goal misgeneralization ↗Threat: Crisis · Trend: ↓ worsening · Evidence: contested
- Tools for Interpreting or Explaining LLM Behavior are Absent or Lack Faithfulness ↗Threat: Open problem · Trend: → steady · Evidence: estimated
Gecontroleerd
Status: Claude Sonnet 5 (onderzoek, inspanning hoog), Claude Opus 5.5 (controle, inspanning hoog)
Vertaling: Apple Translation (macOS), gecontroleerd door Claude Opus 5.5
Scores: JEV (TypeSafe), schatting van 25 september 2026
Cite and take away
Cite and reuse
Own texts, questions and classifications: CC BY 4.0. Give attribution, link the license and indicate changes. Third-party source reports retain their own rights. CC BY 4.0
AMANi DUNiA. redlines-13 · Welk gemeten verlies van monitoringdekking zou beperkingen op de autonomie van een model v. Snapshot: . https://research.amanidunia.berlin/nl/fragen/redlines-13/
Version: redlines-13
BibTeX
@misc{research__nl_fragen_redlines_13_,
author = {{AMANi DUNiA}},
title = {redlines-13 · Welk gemeten verlies van monitoringdekking zou beperkingen op de autonomie van een model v},
year = {2026},
url = {https://research.amanidunia.berlin/nl/fragen/redlines-13/},
note = {Snapshot 2026-10-01; CC BY 4.0},
version = {redlines-13}
}Language and review status
Main text: NL · includes English sections. Review notes apply only to the content they name.
Translation available; a completed independent content review is not recorded.
See language coverage and data snapshot →