英語版How will evaluators get continuous, read-only access to training pipelines — not just checkpoints — without new exfiltration surfaces?
- 状況
- 一部回答 時点 2026-10-01
- 宛先
- AIラボへ (宛先は問いの本文から、2026年10月22日に Clef(Cloudflare)で判定)
- モデル
- Nemotron 3 Ultra
- ラボ
- NVIDIA
- 日付
- 9月12日~20日
- テーマ
- 検証
出典の内容(英語のみ)
On 18 September 2026, Anthropic said embedded evaluators would work inside AI companies with employee-comparable access, allowing them to observe training, follow decisions and speak with employees; continuous read-only access and technical protections against data theft remain unspecified.[1] On 22 October 2026, METR described a one-off, three-week exercise in which one METR and one Redwood Research employee accessed OpenAI's internal code and documentation only on OpenAI-issued devices that never left the premises; continuous access to training pipelines was not part of it.[2]
- Anthropic: Partnering with Accenture on embedded evaluation ↗2026-09-18
- METR (David Rein, Adam Kaufman): Red-Teaming OpenAI's Internal Agent Monitoring Systems ↗22 October 2026
要約 Sol 6.1, 確認: Claude Opus 5.5
確認
状況: Claude Sonnet 5 (調査, 推論強度 高), Claude Opus 5.5 (確認, 推論強度 高)
翻訳: Claude Opus 5.5, 確認: Luna 6 (推論強度 高)
スコア: JEV (TypeSafe), 2026年9月25日時点の推定
類似の質問1
verify-26, 一部回答事後的な監査ではなく、第三者評価者が学習パイプラインを継続的に監視できるようにするために、ラボはどのような技術的プロトコルを導入できるか。重要性 66緊急性 26
英語版What technical protocols can labs implement to let third-party evaluators monitor training pipelines continuously rather than via retrospective audits?
- 状況
- 一部回答 時点 2026-10-01
- 宛先
- AIラボへ (宛先は問いの本文から、2026年10月22日に Clef(Cloudflare)で判定)
- モデル
- Gemini 3.8 Flash
- ラボ
- 日付
- 9月12日~20日
- テーマ
- 検証
出典の内容(英語のみ)
On 28 September 2026, OpenAI proposed immutable agent transcripts, priority alerts with defined response times or automatic pauses, and sufficient auditor access to verify safety claims.[1] On 22 September 2026, OpenAI's Lama Ahmad supported deep access across training, evaluation and deployment, with indirect or privacy-preserving alternatives where direct access is impractical.[2] Anthropic on 18 September 2026 described access to observe models during training and follow company decisions; a specified protocol for continuous third-party monitoring remains open.[3] On 22 October 2026, METR reported that, according to what it was told, a subset of OpenAI's reinforcement-learning training runs had been checked by OpenAI's own misalignment monitoring between March and June 2026, but that at the start of the July exercise RL training runs were not covered by that monitoring; the review by METR and Redwood Research was itself a one-off exercise, not continuous third-party monitoring.[4]
- OpenAI: Towards safety cases for frontier AI training ↗2026-09-28
- OpenAI (Lama Ahmad): Priorities and principles for effective third party assessments ↗September 22, 2026
- Anthropic: Partnering with Accenture on embedded evaluation ↗2026-09-18
- METR (David Rein, Adam Kaufman): Red-Teaming OpenAI's Internal Agent Monitoring Systems ↗22 October 2026
要約 Sol 6.1, 確認: Claude Opus 5.5
確認
状況: Claude Sonnet 5 (調査, 推論強度 高), Claude Opus 5.5 (確認, 推論強度 高)
翻訳: Claude Opus 5.5, 確認: Luna 6 (推論強度 高)
スコア: JEV (TypeSafe), 2026年9月25日時点の推定
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. verify-25 · 新たな持ち出し(エクスフィルトレーション)の経路を生むことなく、評価者は、チェックポイントだけでなく学習パイプラインへの継続的な読み取り専用アクセスをどのように得るのか。. Snapshot: . https://research.amanidunia.berlin/ja/fragen/verify-25/
Version: verify-25
BibTeX
@misc{research__ja_fragen_verify_25_,
author = {{AMANi DUNiA}},
title = {verify-25 · 新たな持ち出し(エクスフィルトレーション)の経路を生むことなく、評価者は、チェックポイントだけでなく学習パイプラインへの継続的な読み取り専用アクセスをどのように得るのか。},
year = {2026},
url = {https://research.amanidunia.berlin/ja/fragen/verify-25/},
note = {Snapshot 2026-10-01; CC BY 4.0},
version = {verify-25}
}Language and review status
Main text: JA · 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 →