#

Key takeaways

• OpenAI says it disrupted a coordinated campaign aimed at extracting protected reasoning from its models.

• The earliest observed activity was July 1, with spikes on July 24 and 25.

• OpenAI says it fully disrupted the related cluster by July 28.

• The figures described attempted extractions, not confirmed successful extractions.

• The source reports no Moroccan involvement or effects.

What OpenAI disclosed

OpenAI says it identified a coordinated campaign aimed at extracting protected reasoning from its models. The company says the earliest observed activity was July 1. It also says activity spiked on July 24 and 25, before the related cluster was fully disrupted by July 28.

The source frames this as a model-distillation campaign. It says the operators manipulated model conversations rather than breaking encryption or gaining direct access to stored user conversations. OpenAI also says independent researchers disclosed related paths, which it investigated and confirmed.

What the activity looked like

OpenAI reported 16,000 requests using a relevant extraction pattern from over 4,000 users during the spikes. It also reported related prompt patterns across more than 15,000 users. A footnote clarifies that these figures describe attempted extractions, not confirmed successful extractions.

The source says the operators tried to replay encrypted reasoning from one conversation in another. That detail matters because it describes a conversational manipulation pattern, not a direct system breach. The report does not say the attempts succeeded.

Attribution and scope

OpenAI attributes a core cluster to individuals associated with Moonshot AI. It also says it is unclear whether all observed activity came from one actor. This attribution is OpenAI's assessment and should not be treated as an adjudicated finding.

The source also says this newly disclosed incident differs from earlier Anthropic reports about other distillation campaigns. It does not provide a broader comparison beyond that distinction. It also does not assert Moroccan involvement or effects.

Response and controls

OpenAI says its response included account restrictions, signup and infrastructure controls, monitoring, reasoning protections, output checks, and coordination with partner services. It also says it shared findings through the Frontier Model Forum and government channels. The company says further work is continuing.

These measures suggest a layered response. They combine access controls, detection, and model-side protections. The source does not give technical detail on how each control worked.

Why this matters

The report shows how model security can involve more than direct system access. It can also involve repeated prompting, conversation manipulation, and attempts to extract protected reasoning. The source emphasizes that the observed figures were attempts, which is important for interpreting the scale.

It also shows the value of coordinated response. OpenAI says it used internal controls, partner coordination, and external sharing channels. The source does not say whether those steps prevented future attempts, only that the related cluster was disrupted.

Morocco relevance

The source reports no Morocco-specific involvement, impact, or program. For readers, the general lesson is that model-protection issues can involve repeated attempts, not only obvious breaches.

Bottom line

OpenAI says it disrupted a coordinated campaign that targeted protected reasoning in its models. The incident began with activity observed on July 1 and peaked in late July. The company says the related cluster was fully disrupted by July 28, while further work continues.