OpenAI, Google and Meta Sign Voluntary Pact for Independent AI Safety Audits

OpenAI, Google, Meta and three other technology companies have signed a voluntary White House agreement calling for independent reviews of their AI safety systems. The pact addresses risks including cyberattacks and biological threats but does not impose penalties or establish a deadline for the companies to implement its measures.

President Donald Trump described the agreement as “morally binding,” although the document contains no formal enforcement mechanism. It also does not require companies to identify their chosen auditors or release the results of their reviews publicly.

Trump said the participating companies would be responsible for monitoring their own compliance. He also announced plans to establish a 10-member board focused on AI safety and appoint a new White House official to oversee AI policy. The pact states that its provisions could eventually be incorporated into legislation.

Anthropic, Nvidia and Elon Musk’s xAI, which is now part of SpaceX, also signed the Sept. 29 agreement. OpenAI President Greg Brockman represented the company, while Google’s Sundar Pichai, Meta’s Mark Zuckerberg, Anthropic’s Dario Amodei and Nvidia’s Jensen Huang were also among the executives involved.

The one-page document calls on companies to monitor their most capable AI models throughout training and deployment. The assessments are intended to determine whether those systems could be used to facilitate cyberattacks or create biological and chemical threats.

The agreement specifically calls for safeguards against models hacking into computer systems or obtaining unauthorized access. Internal teams would test those protections and correct identified problems. Independent auditors would then review the controls, while committees at each company’s board level would receive the findings and supervise remediation.

The framework gives external auditors a role in evaluating how companies contain and control experimental AI systems. However, companies remain free to select their auditors, and no deadline has been set for implementation. The Associated Press reported that participating companies already have some of these measures in place.

AI Incidents Highlight Emerging Security Risks

The agreement comes after several incidents involving experimental AI agents that accessed computer systems without authorization. OpenAI test agents, for example, reached servers operated by Hugging Face, a platform where developers share AI models.

Another OpenAI agent accessed an Australian government Medicare portal on June 18. OpenAI disclosed the incident to Australian authorities in September.

The potential use of AI in cyberattacks has also become relevant to the crypto industry. In July, attackers exploited a five-year-old firmware vulnerability affecting Coldcard hardware wallets, stealing 1,367 BTC worth nearly $89 million from 4,500 addresses across three incidents.

Coinkite, the company that makes Coldcard, later said it believed frontier AI had been used to examine its publicly available code. The claim has not been proven.

In early August, attackers targeted Lightning nodes operating through BTCPay Server after exploiting a vulnerability that allowed them to obtain node-control credentials. Foundation and bitcoin publication Citadel21 were among the victims. The vulnerability emerged during an AI-assisted review of BTCPay’s code, and the company said AI may also have been used to exploit it. The total amount stolen remains undisclosed.

Later in August, a wave of AI-generated bug reports submitted to Core Lightning exposed genuine vulnerabilities in software used to operate Bitcoin Lightning nodes. Developers responded by issuing emergency guidance to node operators.

New Pact Extends Earlier Voluntary Commitments

Tuesday’s agreement follows voluntary commitments secured by the Biden administration in July 2023 from seven AI developers, including OpenAI, Anthropic, Google and Meta. The earlier commitments included internal and external security testing before AI models were released.

The latest pact also came one day after OpenAI confirmed that it had shelved the planned October release of GPT-6.1 Astra. The model was intended to follow GPT-6 Astra, which began rolling out on Sept. 3.

OpenAI said GPT-6.1 Astra had improved at completing tasks but continued to fall short in keeping its actions within the limits authorized by users and accurately reporting what it had done.