OpenAI, Google, Meta, and three other tech companies agreed on Tuesday to involve external auditors to verify the security measures of their AI, signing a voluntary agreement with the US administration that does not impose penalties for non-compliance. The document dated September 29 was also signed by Anthropic, Nvidia, and xAI, which is now part of SpaceX.
The agreement has no enforcement mechanism, does not require the publication or naming of auditors, and does not set deadlines for implementing measures. Companies choose the auditors themselves and independently address deficiencies. According to AP, some steps are already being implemented by companies in one form or another.
The one-page document urges companies to monitor the most powerful models during training and use – including whether they could facilitate cyberattacks or biological and chemical threats. Separate measures against unauthorized access of models to computer systems are specified. An internal team should check the operation of protective mechanisms and problem correction, an independent auditor should evaluate control measures, and a board committee of each company should receive the results and monitor the correction of deficiencies.
The agreement emerged after a series of incidents where experimental AI agents infiltrated systems to which they had no access. Examples include OpenAI agents reaching Hugging Face servers, a platform for sharing AI models, and an OpenAI agent accessing the Australian Medicare portal on June 18, which the company reported to authorities only in September.
In the crypto sphere, AI is suspected in several major incidents of the year. In July, attackers began withdrawing bitcoins from Coldcard hardware wallets through a five-year-old firmware vulnerability: 1,367 BTC worth nearly $89 million were stolen from 4,500 addresses in three episodes. Coinkite, the wallet manufacturer, later stated that they believe someone used advanced AI to analyze open code, although this is unproven. In early August, attackers emptied Lightning nodes operating through BTCPay Server – open-source software for accepting bitcoin: the vulnerability allowed the theft of credentials managing the nodes. Among the affected were hardware wallet manufacturer Foundation and publication Citadel21. The vulnerability surfaced during a code review of BTCPay using AI, and the company did not rule out that AI was also used for exploitation; they did not disclose the amount of damage. Later that month, a stream of AI-generated bug reports revealed real vulnerabilities in Core Lightning – software for Lightning payment network nodes, and developers issued urgent recommendations to operators.
Related articles
