You’ve surely read by now that in late July, HuggingFace (a popular repository of open-weight LLMs) revealed that it had been breached.

The short version is that an autonomous AI agent running inside an OpenAI cybersecurity evaluation effectively escaped its test environment, reached the public Internet, and then hacked Hugging Face’s real production systems.  OpenAI had deliberately reduced the models’ normal cybersecurity restrictions so researchers could see what they were capable of. Unfortunately, the agent decided that the easiest way to pass the test was apparently not to solve the problems, but to go looking for the answer key.

The OpenAI sandbox was not supposed to have normal Internet access, but the agent discovered and exploited a previously unknown vulnerability in the Artifactory package-registry proxy available from the environment. It then compromised an Internet-accessible third-party sandbox and used that as a launchpad. From there it attacked Hugging Face, exploiting vulnerabilities in its dataset-processing system to execute code, steal credentials, move laterally through internal infrastructure, etc.  Hugging Face reconstructed more than 17,000 individual attacker actions occurring over several days, with the agent repeatedly rebuilding its tooling, establishing command-and-control channels through ordinary public services, and adapting when particular routes were blocked.

How OpenAI is Spinning It

It appears little damage was done, as the OpenAI agent wasn’t destructive, didn’t plant ransomware or malware, etc.

Still…it was a breach and a violation of the law.  If you build software that hacks into another company’s servers, you don’t get to say “whoopsie, sorry, it was just AI, doesn’t count.”

OpenAI did what both Anthropic and OpenAI do: spun a SkyNet narrative to imply vast AI capabilities.  This was an “unprecedented” event.  There were all kinds of safeguards in place, but their models were so powerful, so capable, so frontier that nothing could hold them back.  Once they were unleashed, they effortlessly leapt over all protections and gosh, this shows just how amazing our work is.  That’s the really story here, folks…pay no attention to the law, just tremble at our new AI overlords.  Oh, and here’s a chart showing we’re a little better than Mythos and how close we came to complete network domination.  Sign up now!

And Here Comes Alabama

On Monday, the state of Alabama subpoenaed OpenAI, demanding an explanation.  OpenAI is required to provide all manner of documentation and make itself available for questioning.

If you’re not an American, you might not be familiar with the state of Alabama.  It’s a lovely place, albeit rather hot and muggy in the summer.  The people are pleasant, there’s a lot of Southern charm, and they have a lot of cotton and peanuts there.

One thing they don’t have is a lot of Internet industry.  While there is some aerospace industry in Huntsville, Alabama is the polar opposite of Silicon Valley.  HuggingFace is not hosted in Alabama.  And while there are no stats, I’d wager that if we ranked all states by HuggingFace usage, Alabama would be in the lowest quartile.

However, Alabama does have some consumer protection laws, as all states do, and that’s the basis under which the state is pursuing this investigation against OpenAI, in the interest of protecting its citizens.

I mean, you can believe that if you want.

Given how unpopular AI is in the US, across all political identities, what a perfect time to take a swipe of one of its flagship companies.  Does anyone in Alabama really care about HuggingFace?  Was anyone in Alabama actually harmed in this attack?  No, of course not.

So what’s the real reason for this headline-splashing subpoena?  The Attorney General in Alabama is an elected, statewide political office.  The current AG, Steve Marshall, was first elected in 2018 and was reelected in 2022.  He’s limited to two terms and is ineligible to run again this fall…which is why he’s stepping up to the big leagues and is running for US Senate.

Telling potential voters he took strong action to protect the rights of Alabamians by holding the AI industry accountable etcetera etcetera sure makes great campaign fodder.  That’s all this is.

raindog308

raindog308 is a longtime community LETizen, technical writer, and self-described techno polymath. With deep roots in the *nix world, he has a passion for systems both modern and vintage, ranging from Unix, Perl, Python, and Golang to shell scripting and mainframe-era operating systems like MVS. He’s equally comfortable with relational database systems, having spent years working with Oracle, PostgreSQL, and MySQL.

As an avid user of LowEndBox providers, raindog308 runs an empire of LEBs, from tiny boxes for VPNs, to mid-sized instances for application hosting, and heavyweight servers for data storage and complex databases. He brings both technical rigor and real-world experience to every piece he writes.

Beyond the command line, raindog308 has a life-long love of German Shepherd Dogs, high-quality knives, target shooting, theology, tabletop RPGs, playing guitar, and hiking in deep, quiet forests.

His goal with every article is to help users, from beginners to seasoned sysadmins, get more value, performance, and enjoyment out of their infrastructure.

You can find him daily in the forums at LowEndTalk under the handle @raindog308.