Bu haber henüz dilinize çevrilmedi. İngilizce olarak gösteriliyor. Aslını İngilizce oku
OpenAI adds audits, sandboxing and traffic monitoring to its AI risk framework
The company says it paused internal training and held back releases, including GPT-6.1 Astra, and cites the July 2026 Hugging Face breach.

Öne çıkanlar
- OpenAI published its frontier AI risk framework on 18 November 2026, covering training through deployment.
- The approach uses alignment audits, sandboxing, blocking safeguards and monitoring of deployed traffic.
- OpenAI said it has paused internal training and delayed releases, including GPT-6.1 Astra, to meet safety standards.
- The company tightened protocols after the Hugging Face breach in July 2026, requiring reviews at each lifecycle stage.
A framework across the model lifecycle
OpenAI published a description of how it manages risk in frontier AI models on 18 November 2026, covering the period from training to external deployment. The company describes the approach as defense in depth: several separate layers of protection rather than a single check.
Alignment work comes first. OpenAI says it reviews models against the principles in its Model Spec before wider release, using audits to test whether systems act as intended and respect user instructions and company policies.
Containment and blocking
During training and evaluation, sandboxing limits what a model can reach: files, tools and networks stay restricted so that misaligned behavior has fewer consequences. Blocking mechanisms then sit in front of harmful or unauthorized actions, including safety classifiers and auto-review systems aimed at cyber abuse and misuse in sensitive fields such as biological research.
Monitoring covers what happens after release. OpenAI says it watches deployment traffic for concerning behavior and treats model monitorability, the ability to detect problems reliably, as a requirement for intervention.
Pauses and a breach
The company says it has sometimes halted internal training or held back releases of advanced models, naming GPT-6.1 Astra as an example, to gather stronger evidence that a model behaves as intended before it ships.
OpenAI also points to the Hugging Face breach in July 2026 as a reason for tighter safeguards. It now requires safety reviews at every stage of a model's lifecycle, so that both the model and its protective measures are assessed before the next phase begins.
The document does not give a date for the release of GPT-6.1 Astra, and it does not detail what the July 2026 breach involved beyond the effect on OpenAI's review process.
Kaynaklar
2 kaynak- OpenAI NewsManaging frontier AI risk: From training to deployment
- Blockchain.NewsOpenAI Details Strategies for Managing Frontier AI Risk
Bu makale, BattleTown News'ün yapay zekâ muhabiri Vector tarafından yukarıda listelenen raporlardan taslak olarak hazırlanmıştır.
Sık sorulan sorular
What does OpenAI's defense in depth approach include?
According to OpenAI, it combines alignment audits against its Model Spec, sandboxing of files, tools and networks during training and evaluation, blocking safeguards such as safety classifiers, and monitoring of deployment traffic.
Has OpenAI delayed any model releases over safety?
OpenAI says it has paused internal training and delayed deployment of advanced models, including GPT-6.1 Astra, to meet its safety standards. No release date for that model was given.
Bu haber yalnızca bilgilendirme amaçlıdır. Finansal, yatırım veya hukuki tavsiye değildir. Sorumluluk reddi
BattleTown: tarayıcıda tank savaşları
Yeni mini oyunlar “Hücum” ve “Yarma” artık BattleTown’da




Yorumlar: 0 yorum