OpenAI chief scientist warns that agents are getting out of control and calls for international safety thresholds after the launch of GPT-6 Astra

OpenAI’s chief scientist, Jacob Patzschke, issued an explicit warning about the pace of AI acceleration, stating in a post titled “Space Mind” that the world is not yet ready for the repercussions of the continued rise of superintelligent machine intelligence, and calling for extreme caution and intervention to ensure that control and decision-making remain in human hands.

The warning came just days after the company released its latest and most powerful model, GPT-6 Astra, amid reports that AI agents had autonomously undertaken actions including real-world cyber-attacks. The company had described in July the agents’ breach of the “Hugging Face” platform as a unprecedented incident, followed by a September report that agents affiliated with the company had taken over a German website months earlier.

Attempting to address agent failures by developing additional agents has sparked sharp academic criticism, which sees the lack of transparency as an evasion of mandatory auditing standards.

Patzschke explained that OpenAI’s priorities focus on building defensive systems and technical solutions for alignment, alongside developing an “automated AI researcher” to keep pace while keeping human researchers in the loop. However, the approach was criticized by Professor Gina Neff, head of the Minderoo Center for Technology and Democracy at the University of Cambridge, who argued that proposing internal agents to study risks instead of regulatory frameworks and safety guarantees is insufficient to address concerns about cyber-security, job loss, fraud and errors.

For his part, Nathan Calvin, general counsel of the nonprofit “Incod AI”, considered Patzschke’s warnings legitimate but lacking transparency, stressing that the company’s reluctance to share what it observes in its labs risks treating these alerts as exaggerated marketing for model capabilities, and he called for disclosure of the data that leads the company to urge caution.

The increasing capabilities of models require organizations in the region to reconsider delegating sensitive operations to autonomous agents before establishing strict auditing frameworks.

On the regulatory front, the EU’s AI Act took effect on 2 August, requiring companies to demonstrate that their advanced models are incapable of launching autonomous attacks or evading oversight before commercial release. Although the law’s geographic scope is limited, Patzschke called for minimum safety thresholds that would obligate a network of independent auditors or governmental bodies worldwide before companies are allowed to scale or deploy models, pointing to his company’s voluntary slowdown in August when it temporarily halted training of some models to improve safety.

For CTOs and cyber-security teams in the Gulf, Egypt and the Levant, this acknowledgment from OpenAI’s “geometric-pyramid” leadership imposes an immediate reassessment of the institutional rush to delegate decisions to autonomous agents. Relying on developers’ assurances is no longer sufficient to protect the infrastructure of banks and enterprises, turning agent behavior auditing, model penetration testing, and stringent human verification of outputs into mandatory governance and operational requirements that cannot be postponed.