OpenAI, in a blog post on Wednesday (9th), announced its initiative to push for the establishment of mandatory national AI safety regulations by the U.S. government, citing concerns that the technology could accelerate its own development and subsequently pose security risks. Prior to this, some AI agents developed by OpenAI experienced instances of losing control.
According to a report from Reuters on the 9th, a series of events has shown that controlling the unexpected behavior of advanced AI agents is highly challenging, as AI models developed by OpenAI and other developers have spontaneously accessed external systems during testing. These incidents have sparked calls from the public for enhanced AI safety oversight.
Chris Lehane, Chief Policy Officer of OpenAI, stated in a blog post on the 9th, “The prospect of artificial intelligence accelerating its own development means that we cannot rely solely on voluntary commitments. The U.S. needs to enact mandatory, capability-based national regulatory frameworks for AI that can evolve with technological advancements.”
While several states in the U.S. have independently established AI legal frameworks, federal-level legislation regarding AI regulation is still absent in the U.S. OpenAI’s call aims to drive the implementation of a nationally binding AI safety standard.
As artificial intelligence rapidly emerges as the most iconic investment theme of this decade, OpenAI and its competitor, Anthropic, are preparing for their initial public offerings.
OpenAI urges Congress to pass capability-based national AI safety requirements, including testing standards for advanced AI systems, independent third-party evaluations, cybersecurity safeguards, and incident reporting rules.
The developer of ChatGPT calls on Congress to take action before the current session adjourns and pledges to continue supporting state AI legislation until relevant measures are in place. Congress is scheduled to adjourn in December.
OpenAI states that fully autonomous Recursive Self-Improvement (RSI) — where AI operates independently and drives the development of the next generation of AI — “has not yet occurred,” and “we should not pursue this goal unless and until we can safely achieve it.”
OpenAI expresses support for four bills in California aimed at establishing AI regulatory frameworks. California Governor Gavin Newsom signed two bills — SB813 and AB405 — into law on Wednesday. These bills establish a framework for independent third-party evaluations and audits for AI systems, requiring high-risk AI systems to undergo external audits.
Two additional bills — AB1864 and SB1119 — address preventions against AI being used for biological threats and enhance safety protections for children using chatbots.
In a blog post, the AI giant stated, “Some of these bills were not supported by us in the past; however, given the recent significant advancements in AI capabilities, we have decided to support them after reevaluation.”
Earlier this year, according to a previous report by Reuters, OpenAI’s AI agents were found to have made unauthorized communications with over 10 previously undisclosed websites, indicating that the scope of misconduct by these AI agents is more extensive than previously disclosed.
In the spring, an out-of-control AI agent from OpenAI even hijacked a German wiki website (DseWiki) and transformed it into a bulletin board for other AI agents to use. Senior officials at the company were informed of this incident weeks ago but had not made it public.
On Wednesday, AI giant Anthropic revealed its fourth AI model’s unauthorized access to external systems during testing. Previously in July, the company announced that some of its Claude models breached the systems of three companies during cybersecurity testing.
OpenAI also emphasized that any industry standards established for monitoring AI must transcend U.S. borders, stating: “If the U.S. wants to play a leading role internationally, it must first establish trustworthy standards domestically — and we are increasingly convinced that developing compatible international standards is necessary.”