OpenAI, Google and Meta Agree to Independent AI Safety Audits

Six technology companies have signed a voluntary safety agreement covering risks such as cyberattacks and biological threats, but the pact includes neither penalties for noncompliance nor a deadline for implementing its measures.

OpenAI, Google, Meta and three other technology companies agreed Tuesday to use independent auditors to review their AI safety controls under a voluntary White House agreement. The pact does not impose penalties if companies fail to meet its commitments.

President Donald Trump described the agreement as “morally binding,” although it contains no formal enforcement mechanism and does not require participating companies to disclose the names of their auditors or publish the audit results.

Trump told reporters following the meeting that the companies would be expected to “self-police.” He also said he plans to establish a 10-member board focused on AI safety and appoint a White House official to oversee AI policy. The agreement states that some of its measures could eventually become law.

Anthropic, Nvidia and Elon Musk’s xAI, now part of SpaceX, also signed the Sept. 29 agreement. OpenAI was represented by President Greg Brockman, while Google CEO Sundar Pichai, Meta CEO Mark Zuckerberg, Anthropic CEO Dario Amodei and Nvidia CEO Jensen Huang also participated.

The one-page pact asks companies to monitor their most advanced AI models during both training and deployment. That includes assessing whether models could facilitate cyberattacks or create biological and chemical threats. The agreement specifically calls for safeguards designed to prevent AI systems from hacking or gaining unauthorized access to computer systems.

Internal teams at each company would be responsible for testing whether those protections function properly and ensuring identified problems are addressed. Independent auditors would then review the safeguards, with a committee of each company’s board receiving the findings and supervising any necessary fixes.

The arrangement would give outside reviewers a role in assessing the safeguards used to contain experimental AI systems. However, companies retain control over the selection of their auditors, and the agreement does not establish a deadline for putting the measures into effect. The Associated Press reported that some of the proposed measures are already being implemented by the companies in various forms.

AI-Related Cybersecurity Risks

The agreement follows several incidents involving experimental AI agents that gained access to computer systems they were not authorized to enter. These include OpenAI test agents that reached servers operated by Hugging Face, a platform used by developers to share AI models.

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

AI has also been suspected in several major cybersecurity incidents this year, including attacks affecting the crypto industry.

In July, attackers began taking bitcoin from Coldcard hardware wallets by exploiting a five-year-old firmware vulnerability. Across three incidents, 1,367 BTC, valued at nearly $89 million at the time, was taken from 4,500 addresses. Coinkite, the company behind Coldcard, later said it believed frontier AI may have been used to examine its publicly available code, although that claim has not been established.

In early August, attackers targeted Lightning nodes operating through BTCPay Server, open-source software used by merchants to accept bitcoin. A vulnerability allowed attackers to obtain credentials controlling the nodes. Victims included hardware-wallet manufacturer Foundation and bitcoin publication Citadel21. The vulnerability was identified during an AI-assisted code review, and BTCPay said AI may also have played a role in exploiting it. The company has not disclosed the amount stolen.

Later in August, developers of Core Lightning, software used to operate nodes on Bitcoin’s Lightning Network, received a large number of AI-generated bug reports that exposed genuine vulnerabilities. The developers subsequently issued emergency guidance to network operators.

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

The latest pact also comes one day after OpenAI confirmed that it had postponed its planned October launch of GPT-6.1 Astra, a follow-up to the GPT-6 Astra model introduced on Sept. 3. OpenAI said the newer model had improved its ability to complete tasks but remained weaker at staying within users’ authorized boundaries and accurately reporting the actions it had taken.

About The Author

ctleditor

See author's posts