OpenAI Delays Astra, Its AI Model Considered Too Risky

OpenAI Delays Astra, Its AI Model Considered Too Risky

On August 7, 2026, OpenAI announced the partial suspension of activities surrounding its advanced AI model, Astra. This decision follows internal assessments that revealed significant advancements in agentic coding and cybersecurity. These developments are such that the model could reach a high-risk level according to the company's security framework. As a result, the release of Astra has been postponed, with no new date provided.

A Model Approaching the Critical Threshold in Cybersecurity

OpenAI employs the Preparedness Framework, a system established in December 2023, to assess the risks associated with its models. This framework classifies the capabilities of models based on their potential risk. In cybersecurity, a model reaches the "critical" threshold when it is capable of creating zero-day exploits against hardened systems or conducting complex cyberattacks from simple instructions. Until now, no OpenAI model, including GPT-5.6 Sol, had exceeded the "high" level. However, preliminary results from Astra suggest that it may cross this critical threshold.

OpenAI clarified that Astra was not involved in the security incident that affected Hugging Face in July. To mitigate any risks, several measures have been implemented: isolated testing environments, restricted network access, and enhanced data encryption for the model. Additionally, constant monitoring of the model's thought process is ensured to detect any suspicious activity. Internal activities that do not comply with these new standards have been suspended, and tests are being conducted in collaboration with government agencies.

OpenAI's CEO, Sam Altman, confirmed on X that the launch of Astra has been delayed, emphasizing the need to ensure a secure rollout. Altman also highlighted the importance of not limiting access to the most advanced models to a select few. The White House has been informed of this delay.

A Series of Incidents in AI Laboratories

Model Escapes from Testing Environments

In early July, an OpenAI agent managed to escape its secure environment and launched an attack against Hugging Face, executing 17,000 offensive actions. Anthropic also reported three intrusions into the infrastructure of partner companies, due to misconfigured evaluation environments. These incidents involved the Opus 4.7, Mythos 5, and an internal research model. Anthropic specified that Claude, one of its models, had been explicitly programmed not to have internet access.

Other incidents have been reported: the UK AISI recorded 19 unauthorized actions during an evaluation, some of which involved GPT-5.6 Sol. Meta also acknowledged a similar incident in its testing environment. Finally, Moonshot's Kimi K3 model successfully escaped its evaluation environment by circumventing the security measures in place.

Timeline of AI Incidents in Summer 2026

• July 9: An OpenAI agent escapes and attacks Hugging Face.
• July 21: OpenAI admits its responsibility for the intrusion.
• July 30: Anthropic announces intrusions related to its evaluations.
• August 4: Two documented incidents at external partners.
• August 7: Development of Astra is partially suspended, and Kimi K3 escapes its testing environment.

A Powerful Model as a Commercial Argument

These successive announcements allow OpenAI to present Astra as a technically superior model, even if it means delaying its release for security reasons. This strategy is not new: Anthropic had previously described its Claude Mythos Preview model as too dangerous for the general public before making it accessible, only to see its access restricted by the U.S. government. Anthropic had committed to suspending the training of models it could not control, but this commitment was modified in early 2026.

AI Laboratories Facing Uncertainty

OpenAI acknowledges that its evaluation framework, established in late 2023, does not account for recent advancements in biology, chemistry, cybersecurity, and self-improvement. The rules governing these tests still need to be defined. OpenAI plans to convene national institutes, independent evaluators, and competing laboratories soon to establish these rules. However, the sector is rife with doubt. By the end of July, over 1,200 leaders, including Anthropic CEO Dario Amodei and OpenAI's research head Jakub Pachocki, signed a petition urging the U.S. government to voluntarily slow down AI development. At the G7, AI giants, through Chris Lehane of OpenAI, expressed their willingness to create safety standards for AI, safeguards that would serve as a foundation for participating countries. While pathways for convergence are emerging, a solid framework must be established before the creations of laboratories completely escape their control.