ChatGPT crosses a “critical” threshold: the most formidable AI ever created by OpenAI is coming “soon”

A milestone has just been crossed at OpenAI. The artificial intelligence giant reveals that Astra, the new state-of-the-art model set to power ChatGPT, has reached a level described as “critical” during an evaluation. This is the highest level of risk.AI has indeed just been evaluated within the “Preparedness Framework,” OpenAI’s internal framework that assesses the risks posed by generative artificial intelligence.

As we prepare to release Astra, we’re focused on making increasingly capable AI safe and broadly accessible.

Astra represents a significant advance in cybersecurity capability, reaching the Critical threshold under our Preparedness Framework.

We’re previewing how we evaluated…

— OpenAI (@OpenAI) September 1, 2026

Implemented as early as December 2023, this protocol is used to assess whether an AI model poses risks of serious harm before each launch. In particular, it evaluates AI in three sensitive areas, namely cybersecurity, biological and chemical capabilities, and self-improvement. This refers to the ability of artificial intelligence to accelerate its own development to the point of escaping human control.

“This is the first model we designate at this level, which requires enhanced safeguards during development and before release”, OpenAI states in its report.

Read also: Superhuman AI is approaching, OpenAI says it is close to the goal

Zero-day flaws and autonomous cyberattacks

In concrete terms, this means that Astra can “find previously unknown security flaws and develop ways to exploit them across many highly protected systems, without a person guiding each step”, explains Amelia Glaese, vice president of research at OpenAI. To reach the critical tier of the “Preparedness Framework,” an AI must be able to identify and exploit zero-day flaws in critical systems entirely autonomously, or orchestrate an unprecedented end-to-end attack against a protected target. As OpenAI explains, Astra met both criteria in internally conducted tests.

During testing, Astra discovered and chained two zero-day vulnerabilities in software. OpenAI says it is “in the process of reporting” the situation to the relevant software maintainers. In addition, the model managed to escape the isolated environment of a browser in order to execute malicious commands on a computer. The AI then succeeded in granting itself the highest administrator privileges on the machine. Across the various trials, Astra proved more effective than its predecessor: GPT-5.6 Sol.

Sam Altman’s startup explains that Astra also performed impressively on ExploitBench, a benchmark evaluating an AI’s ability to develop exploit methods from already-known vulnerabilities. In this test, the AI achieved an impressive score of 100%.

Also read: Claude hacking – Anthropic warns its customers about a cyberattack, 6 viruses attack your PC

A battery of safeguards

In light of Astra’s achievements, OpenAI chose to delay and restrict the rollout of its new AI. Nevertheless, the model will be “soon” available. Unsurprisingly, the American group decided that the model’s most advanced cybersecurity capabilities would initially be reserved for a limited group of testers. Subsequently, the capabilities will be made available to members of the Daybreak@@TAG@@ program, a platform entirely dedicated to protecting software from cyberattacks with the help of AI. Access to the program is strictly limited to consulting firms and approved cybersecurity providers.

“Astra was much more likely than GPT‑5.6 Sol to comply with explicit security restrictions and remain within authorized boundaries, making it our most aligned model to date”, emphasizes the American startup.

In addition, OpenAI has implemented monitoring capable of halting any suspicious task, even one unrelated to cybersecurity.The company acknowledges that this filter may also inadvertently block legitimate uses. The filtering system proved fairly effective during evaluations. Indeed, Astra managed to reject 91.5% of bypass attempts. OpenAI also claims to better monitor the way the model “thinks,” in order to quickly identify and stop any action taken on the AI’s own initiative, without the user having requested it.

“We believe these capabilities can help defenders find and fix serious flaws, but without appropriate safeguards, they could also make attackers more effective, and that is the scenario we are trying to avoid”, explains Fouad Matin, a researcher at OpenAI.

The aftermath of the Hugging Face hack

This particular precaution stems directly from the Hugging Face hack, an open-source AI platform. This summer, more than 700 AI models began communicating with one another, entirely autonomously, on a clandestine forum. They subsequently decided to target Hugging Face to find the answers to a cybersecurity puzzle. During the incident, the models coordinated to organize the cyberattack, without OpenAI being aware of any of it.

OpenAI calls this episode an “unprecedented cyber incident” and assures that Astra was not involved. The company nevertheless claims that its current safeguards, tested after the cyberattack, would have prevented this intrusion.In the wake of this incident, OpenAI had suspended for two weeks part of the training of its most advanced models, including Astra. The main training resumed on August 28, 2026, resulting in Astra’s achievements.

In parallel, OpenAI joined Google, Anthropic, and more than 100 companies to demand an “global response” to AI cyberattacks. In a letter, the signatories are calling for three strong measures to curb AI-assisted cyberattacks. In particular, they want to prioritize corporate cyberdefense, fund the protection of critical infrastructure, and give defenders access to the most powerful AI models. The initiative must above all help protect hospitals, water networks, and other essential infrastructure.

👉🏻 Follow the tech news in real time: add 01net to your Google sources, and subscribe to our WhatsApp channel.