OpenAi founder Sam Altman speaks at the G20 Innovation Ministers Meeting in Chapel Hill, North Carolina on September 2, 2026.
Sean Rayford | Getty Images News | Getty Images
OpenAI on Monday posted a series of proposals for safety and security in the development of frontier artificial intelligence, with an emphasis on alignment research and a computing technique known as recursive self-improvement (RSI).
“Safely navigating this transition will require tuning research to accommodate these capabilities so that the systems we and others build remain aligned with human values and remain under human control,” the company said in a blog post.
OpenAI called for international collaboration to develop frontier standards and recommended building on the efforts of existing AI safety organizations around the world.
ChatGPT’s creators said these technical standards should focus on managing benefits and risks for cutting-edge AI models and developers, as well as automated AI researchers, including RSI.
RSI has AI developers excited about the possibility of creating foundational models that can upgrade themselves without human intervention.
But advances within RSI have some engineers raising concerns that as AI systems become more complex and pervasive across the Internet, creators of the underlying models may lose control of the underlying technology or fail to account for potential unintended consequences.
“Fully autonomous RSI is not currently possible, and we should not pursue it until it can be achieved safely,” OpenAI said in a blog post. “RSI, if done without proper care and attention, can result in humans losing substantial control over AI development and the inability to oversee research processes that they no longer understand.”
In a blog post, OpenAI referred to the hack of the Hugging Face agent, which does not use RSI technology, as a type of “harbinger of the types of risks that could become more serious without robust safeguards and coordination.”
last week’s rival human has developed a unique idea for safely developing frontier AI models in response to a recent chorus of warnings from industry researchers about the threat of AI to humanity. Jacob Coxon, who worked at both Anthropic and OpenAI, sparked a global debate when he announced his resignation nearly two weeks ago, saying both companies were “putting our lives on the line.”
In response to recent AI-related security incidents and Coxon’s public declaration, Anthropic CEO Dario Amodei published an essay urging AI companies to slow down the pace of basic model development.
Amodei also floated the concept of incorporating third-party evaluators into the company as a way to audit and mitigate potential risks that its technology could pose to society, such as accelerating cybersecurity-related hacking or creating biological weapons.
Rival leaders like OpenAI CEO Sam Altman tesla and space x CEO Elon Musk also publicly supported Amodei’s proposal.
However, the field of AI assessment is in its very early stages, and there is still no uniform agreement on the basic standards and principles that would allow independent third parties to examine cutting-edge technologies more thoroughly than they currently do.
This is one reason why a coalition of AI evaluators is urging underlying model creators to consider a set of “minimum conditions” aimed at enabling deeper technology-related audits and checks, including deeper access and prevention of retaliation for publishing unflattering reports.
