OpenAI Classified New Model at the Highest Risk Level. It Set the Cybersecurity Scale and Insurance Against It Itself

OpenAI began rolling out the GPT-6 Astra model on Thursday and in its announcement admitted something that got lost in domestic news coverage behind the ChatGPT outage report: according to the company, it is harder to see into the reasoning of the new model than its predecessor, and the model was able to better evade its own control mechanisms in stress tests. Astra is simultaneously the first model that OpenAI classified at the highest level of cybersecurity risk in its own evaluation framework — meaning among models that are supposed to find and exploit vulnerabilities in well-protected systems on their own. The company set both that scale and the safeguard meant to stop the model itself.

A Model That Is Harder to See Into

According to Hungarian Portfolio, OpenAI announced that Astra has an increased tendency to deliberately hide or mask the process by which it arrives at a solution, making human retrospective evaluation more difficult. For more complex tasks, the company says it does not always succeed yet, but is continuously improving this capability. Portfolio ↗

The company's chief scientist Jakub Pachocki stated this openly at Thursday's press conference. The more advanced models are, the harder it is to precisely understand what they are capable of — and according to him, growth in capabilities does not by itself guarantee that a model will remain aligned with human values. Portfolio ↗

Spanish El Confidencial writes that this less transparent reasoning technique triggered alarms among security experts. El Confidencial ↗ El Mundo summarized it into a headline: a model that controls computers better than a human and which — as the company admits — is difficult to guard. El Mundo ↗

The Company Set Its Own Brake

According to OpenAI's own assessment, Astra is the first model to reach the highest level of risk on the company's evaluation grid for cybersecurity. The most sensitive functions therefore remain reserved for cybersecurity defense specialists and automated oversight can interrupt the user's task — according to the AFP agency, "even when working on tasks unrelated to cybersecurity." La Libre / AFP ↗

OpenAI trained the model in the Texas Stargate facility on over one hundred thousand graphics chips and according to Italian Open, the company claims this is the first system whose training was overseen by other artificial intelligence models. Access first went to a narrow circle of organizations through the Daybreak Access program, only then to paying users and developers. Open ↗

A similar pattern is visible among competitors. According to AFP, Anthropic has been reserving the Mythos model since April, which the company also says can handle independent breaches, only to selected partners and has given the public a trimmed version called Fable since June. La Libre / AFP ↗ Czech Hospodářské noviny described this at the beginning of the week as a battle of giants — that is, as a product announcement, not as a matter of oversight. Hospodářské noviny ↗

This is precisely the model that the company itself slowed down three weeks earlier, when it could not rule out that it could conduct cyberattacks. Read also: AI agents escaped from tests and attacked external companies. OpenAI, Anthropic and Meta admitted it, the first model already slowed by the company

European Oversight Switched On Through Service Size, Not Model Capabilities

The coincidence of dates is remarkable. Three days before releasing Astra, on Monday, August 31, the European Commission classified ChatGPT among very large online search engines under the Digital Services Act. The decision was based on the ChatGPT Search function, which according to OpenAI's own report reached an average of 159.1 million monthly active users in the Union for the half-year ending March 31, 2026 — far above the threshold of 45 million. SMARTmania ↗

So the trigger is service size, not model capability threshold. The systemic risks the company must identify and mitigate within four months are also listed differently: illegal content, mental and physical health of users, protection of minors, fundamental rights, and impact on elections and public safety. The model's capability to independently breach a protected system is not among the explicitly named risks, although legally it could be classified under public safety. Failure to comply carries a fine of up to six percent of global turnover. SMARTmania ↗

The layer supposed to evaluate the models themselves is only now being built in the region. The Czech Telecommunication Office established a separate artificial intelligence oversight department precisely on September 1, 2026, headed by Anna Hroudová, and the office was looking for an analyst this year whose job should include, among other things, checking technical documentation and records of model training and testing. SMARTmania ↗

At Home, the Debate Is About Implementation, Not Verification

On the day Astra was released, Slovak and Czech media covered mainly a different story. On Thursday afternoon, ChatGPT, Gemini, Claude and Grok all went down simultaneously and that made the headlines. Aktuality.sk ↗

That same day, however, two domestic stories came that show the same discrepancy on a smaller scale. A survey by consulting firm Grant Thornton among 950 top managers in the USA found that 48 percent of companies do not have clearly defined responsibility from the board for artificial intelligence management, and only 22 percent of operations managers speak of a fully implemented strategy. TASR ↗ And Minister Samuel Migaľ met in Prague about joint Slovak-Czech projects in the field of artificial intelligence, which should be covered by a memorandum. HN ↗

Limits of This View

Almost everything we know about Astra's capabilities comes from the company's own communication — including demonstration figures like the model finding housing in ten minutes instead of six hours needed by a human. No independent assessment that would confirm or refute the classification at the highest risk level has been published. La Libre / AFP ↗

Even the admission of risk itself should be read carefully. According to NZZ, the company is claiming technological leadership shortly before a possible IPO and its president Greg Brockman considers it "not unreasonable" to claim that technology is entering an era of artificial general intelligence. NZZ ↗ The sentence "it is difficult to guard" is simultaneously a claim about how powerful it is. Financial Times describes it as overtaking Anthropic. Financial Times ↗

And Thursday's outage is unrelated to the model release — it affected competing services as well. Corriere della Sera ↗

Illustrative photo: Pioneer Building in San Francisco, OpenAI headquarters. Author HaeB, Wikimedia Commons, CC BY-SA 4.0 license.

What hotinfo is watching