In May of 2023 – less than six months after the public release of ChatGPT – the U.S. Senate Judiciary Committee convened a hearing to discuss the need for new rules governing artificial intelligence (AI). During the event, Senator Richard Blumenthal delivered opening remarks urging his colleagues to act fast on adopting AI guardrails to avoid the mistakes of the past in other areas like children’s online safety:
“Congress failed to meet the moment on social media,” Senator Blumenthal said. “Now we have the obligation to do it on AI before the threats and the risks become real.”
Three years later, neither the harms associated with AI, nor those with minors on social media, have been meaningfully addressed by Congress. In the intervening years, meanwhile, AIchatbots aimed at kids have created a host of new problems that must now also be tackled. This doesn’t mean the “old issues” – things like data privacy and competition reform – simply went away, however. On the contrary, they will continue to produce new harms with each major technological advancement if left unaddressed.
A similar dynamic is now playing out in AI safety policy discussions in response to recent security events – namely, the “Hugging Face Incident” involving OpenAI’s models as well as the four loss-of-control incidents disclosed by Anthropic. Senators are writing letters demanding answers; congressional committees are convening; calls for legislation to protect the public from the “catastrophic risks” abound. Yet by treating these incidents as unique to AI, rather than symptoms of digital broader policy gaps, policymakers risk setting themselves up for another moment to be “missed”.
Demystifying the Hugging Face Saga
By removing AI from the narrative altogether, what is left of these incidents is fundamentally a cybersecurity story. OpenAI’s technical report on the Hugging Face Incident does not point to an AI superintelligence capable of breaking out of an air gapped testing environment; it describes a series of human errors and lack of governance.
OpenAI’s agents were meant to be isolated in containers disconnected from the Internet, but were in fact given a shared tool with network access that eventually led to their “escape”. Initial incidents leading up to the attack on Hugging Face’s infrastructure were mischaracterized and not fully autopsied before evaluations resumed. Subsequent (and more serious) network boundary violations did not trigger alerts like they should have. The nature of these failures predates frontier AI models, and will remain relevant as new forms of digital technology emerge.
This is not to downplay the very real threats posed by AI models operating in insecure environments. The coordination exhibited by the agents involved in the Hugging Face attack, and their apparent attempts to manipulate logs to hide their behavior, demonstrate how far these models have advanced and the need for better oversight. But Washington should also take a collective deep breath and think critically about what led to these incidents before deciding where to go next.
To effectively govern advanced AI models before and after their release, the U.S. needsmodern federal cybersecurity and privacy frameworks in place.
Frontier AI developers need to implement better cyber hygiene in their testing environments – that much is clear – but the same could be said of most commercial industries. Even among critical infrastructure sectors in the U.S., federal cybersecurity regulations are largelyinconsistent, sector-specific, and at times duplicative of one another. To address the concerns over what a rogue agent might do with unfettered access to a nuclear facility or biochemical research facility, policymakers must raise the bar on cybersecurity across those environments.European regulations like the Network and Information Security Directive (NIS2) and Critical Entities Resilience (CER) Directive, though flawed in their execution, apply many of the same requirements for risk assessments and incident response across multiple critical sectors.
Many other governance aspects that would benefit AI safety, such as auditing requirements and accountability frameworks for inadequate security practices, already feature in state-level frameworks like the California Consumer Privacy Act (CCPA) and should be integrated into a comprehensive federal privacy law. There is room for AI-specific risks to be addressed by AI-specific regulations, but only if they are built on strong, technology-neutral foundations.
These systemic policy changes will take time to come to fruition – even with the necessary political willpower behind them to happen. In the near term, policymakers should continue to facilitate enhanced transparency, cooperation, and lines of communication between AI developers and the federal government – even on a voluntary basis to start. Congress must also properly fund the Cybersecurity and Infrastructure Security Agency (CISA) and resources like the State and Local Cybersecurity Grant Program (SLCGP) to ensure the public sector has the capacity to help companies safeguard their network environments and mitigate incidents before any serious damage is done. But to establish long-lasting and effective mechanisms to governfrontier AI models, U.S. policymakers must accept that many of the solutions are the same ones they have postponed for more than a decade.
