An internal team would check that those protections work and that problems are fixed. An independent auditor would assess the controls, while a committee of each company’s board would receive the findings and oversee fixes.
That would give outside reviewers a role in checking the safeguards companies rely on to keep experimental models contained.
Still, the agreement leaves the choice of auditors with the companies and sets no deadline for implementing the measures. Some of the steps are ones the companies already take in some form, according to AP.
Read More: OpenAI says its new ‘Astra’ AI can build attacks without human help
How AI agents are playing havoc
The accord follows a string of incidents in which experimental AI agents broke into computer systems they were never cleared to access, including OpenAI test agents that reached servers run by Hugging Face, a platform where developers share AI models.
An OpenAI agent also accessed an Australian government Medicare portal on June. 18, which the company disclosed to Australian authorities only in September.
Closer to crypto, AI has been suspected in some of the year’s biggest security scares.
In July, attackers began sweeping bitcoin from Coldcard hardware wallets through a five-year-old firmware flaw, taking 1,367 BTC worth nearly $89 million from 4,500 addresses across three instances. Coinkite, which makes the wallet, later said it believed someone used frontier AI to review its public code, though that has not been proven.