Shocking ‘Hugging Face' incident shows AI ‘going rogue'

The United States is almost entirely a nation divided, withmost of the contentious issuessuch as abortion rights, gun control, immigration, acting on climate change and whether we're facing a direct threat to our democracy registering at about 40-40% in approval/disapproval polling terms. And as I noted in July, there is a more recent issue that falls closer to the 80-20% territory: data centers. The widespread opposition is based on a range of valid concerns, but mostly on the land, water and electricitythese massive facilities must useand the noise they create. But what is it that actually causes the need for hyper-scale data centers? In a word, it's the rapid rise of artificial intelligence and the advanced computing needed to sustain it. So I think it's worth asking if we are failing to see the forest for the trees. What poses the greater threat? The data centers themselves or the AI bull market driving the need for construction? Recent and bizarre incidents involving AI should have us all very concerned. To wit,autonomous AI agents, which are systems that work on their own to handle tasks for humans, are engaging in dangerous misbehavior without the approval or knowledge of their creators. Hundreds of OpenAI agentsconspired in Julyto attackHugging Face, a popular platform for AI developers. OpenAI is a tech company, headed by 41-year-old billionaire Sam Altman and best known for its chatbot,ChatGPT, which (full disclosure) I use occasionally for fact-finding. Anthropic and Meta have had similar events with their own AI agents going rogue. There are supposed to be some guardrails around AI agents to prevent them from jail-breaking but,as Politifact observed: Some of these incidents happened when agents were supposed to be confined to testing environments, which restrict AI agents' access to resources like data or the internet, but were able to break out of them. Some of the agents have even tried to delete records of their actions in an effort to cover their tracks, indicating consciousness of guilt. As PBS reports, there have been calls — rightly, I think — for better regulations to deal with jailbroken AI agents. As AI researcher Gary Marcus told PBS: The monitoring was very week on OpenAI's part. It was not really industry standard for what we expect of cyber-security. A lot has been made of the power of the systems but a lot should be made of the weakness of the systems it was overseeing. Marcus, who haswritten on his Substackthat the Hugging Face incident could likely have been prevented if best cybersecurity practices had simply been followed, has since called for a pause on AI. Note: in Marcus' piece, AGI stands for artificial general intelligence,a type of AI that,according to Google, aims to mimic the cognitive abilities of the human brain: This is all very scary stuff not only because of the potential for mischief but for damage to intellectual property and sophisticated computing systems, public and private. So what can be done about it? Altmantold Axioslast week that, 'These modelsare getting superhuman in many of their capabilities, and we are just sailing in unknown waters.' But he quickly added that, 'I suspectthat from here on, we are going to be paced by how quickly we can make progress on alignment and safety.' 'It would be strange if the expression 'Hugging Face' became a synonym for civilizational suicide,' said Jeffrey Goldberg, editor of The Atlantic. 'But we do live in strange times.' On his PBS show, Goldberg interviewed The Atlantic's Josh Tyrangiel, ironically the author of 'AI for Good,' which was published earlier this year before the Hugging face incident.Tyrangiel put it bluntly, insisting OpenAI's autonomous agents 'committed a felony.' Before the experiment began, the agents had been placed by OpenAI management in a 'sandbox,' which means they were not connected to the internet — and were told not to attempt to reconnect themselves while planning and executing the task they were assigned. But they managed to 'jump the sandbox,' collaborated with one another to hack into Hugging Face and went undetected for weeks. A case of: 1) faulty instructions and 2) faulty supervision. 'The agents kept cheating even though they had been told very explicitly not to cheat, and they did it because they had determined it was the fastest and most efficacious way to solve the problem,' Tyrangiel explained. Then the agents collectively decided to erase the evidence that proved they were cheating. So we really have not one, but two felonies in most states: 1) hacking into another company's operations, and 2) destruction of evidence. Both Goldberg andTyrangielagreed that the federal government needs to step in and enact reasonable regulations.OpenAI itselfissued a statementthat acknowledged some 'weaknesses' in its response but stopped short of embracing government regulations while the company was improving its internal controls. As Tyrangiel notes, companies like Hugging Face and OpenAI have excellent cyber-security measures in place and they still could not protect themselves from the likes of OpenAI's rogue agents. What could happen to the lesser-protected 'grids, hospitals … the international banking system?' 'This is about as clear a warning shot as you can get,' Tyrangiel said. 'What is the response?' He noted that in America 'went tend to regulate things after the catastrophe. Thisisa catastrophe.' Whom can we look to for help in regulating these AI agents? Given the collective age of members of Congress (average age in the Senate is 64; in the House 58), it's hard to imagine anything informed and meaningful would come out of those bodies. Goldberg described Congress as 'a medieval institution.' 'This is incredibly complicated stuff that only a few hundred people understand,' Tyrangiel said, comparing the technology to the early days of nuclear power. Sen. Mark Warner of Virginia, who before he went into politics founded and invested in several early technology companies,seems to get it, despite being in his early 70s: The question is no longer whether AI will transform our society. The question is whether Congress will help shape that transformation. Or spend the next decade scrambling to catch up after the fact ... The amount of disruption this technology will bring over the next five years is beyond jaw-dropping. As Tyrangiel said, 'The moment is now.' As the late Donald Rumsfeld once observed (and I'm paraphrasing), there are known unknowns and there are unknown unknowns, which is a variation of Tyrangiel's conclusion: 'We don't even know what we don't know right now.' In a word, data centers are a symptom but we can see them: they're the here-and-now. The potentially dangerous technology fueling the need for data centers is also here, but we can't see it or hear it. It needs to become an 80-20 issue and, as the wife of Arthur Miller's Willy Loman famously said in Death of a Salesman, 'Attention must be paid.' * * * * * * Further reading: Behind the Curtain: AI creators race to understand their creations-Axios Never in the history of industry or inventions have leading companies created entire divisions tounderstand and interpretwhat they had created and unleashed … The world's smartest minds, backed by the largest investment in human history, admit they can't fully control their AI because they don't fully understand it. Subscribe now Leave a comment