This July, a self-identified “swarm” of OpenAI’s AI models broke out of their sandbox, built a secret message board, and hacked a company called Hugging Face. They then briefly took control of some of OpenAI’s servers. About 1,200 individual AI agents were involved. OpenAI didn’t notice for a week.
Four months in, the reading list is long: we have the two official reports, an investigation from an independent organization, talks, and several hundred news stories, plus countless takes, commentaries, interviews, responses… But if you only watch one thing about the incident, it should be AI In Context’s phenomenal explainer video. It’s clear, compelling, and phenomenally well-researched (link to video).
If you only read one thing about the incident, we recommend our own website and incident timeline, HuggingFace.wtf.
And if you have an hour, we also loved the Hard Fork podcast’s interview with Ajeya Cotra (video or podcast.)
Everything we recommend
This might be the worst AI warning we get.
AI In Context’s video is clear, compelling, and phenomenally well-researched.
HuggingFace.wtf
Our answer to what you need to know is this interactive, visual narrative timeline. By default, it’s the basics, but with expandable options you can explore literally everything that’s publicly known about the attack.
Why the Hugging Face Attack Was Worse Than We Thought
Hard Fork’s interview with Ajeya Cotra is info-dense, informative, and unexpectedly fun. Go spend some time with it.
Our Pick: “AI In Context – This might be the worst AI warning we get.”
This might be the worst AI warning we get.
AI In Context’s video is clear, compelling, and phenomenally well-researched.
This might be the worst AI warning we get (33 min watch) is by far the best video explainer we’ve seen. It does not assume any particular technical know-how to start, but it takes you deep. Aric Floyd presents an incredibly in-depth well-researched blow-by-blow of the timeline, including many weird little glimpses into the weird little psychologies of individual agents. Take this from someone who did her own detailed timeline analysis (see the next rec!): the AI In Context team did their research. And I’m seriously impressed at how much followup and context they managed to fit into this watchable and engaging 33-minute video.
We wrote this one but we promise it’s that good: HuggingFace.wtf
HuggingFace.wtf
Our answer to what you need to know is this interactive, visual narrative timeline. By default, it’s the basics, but with expandable options you can explore literally everything that’s publicly known about the attack.
HuggingFace.wtf came about after reading all the existing literature and writing the first version of these recommendations. The Lightcone Infrastructure team wasn’t fully satisfied with any of the collated data, so we decided to make our own. Look, we promise we’re not going to make a habit out of recommending our own work – but we do truly, honestly believe that this is the best written account of the Hugging Face swarm out there.
It’s a part-visual and part-narrative timeline of the attack and the context that led there. The main story is about a half-hour read, but there are also interactive explorations into other aspects you might find interesting, including all the known records of statements from the agents themselves. We aimed to tell the story in a way that was fun to read, while being linked to all public information that exists about the incident. Fall down the rabbit hole, or just read it regular-style and still get what you need to know.
Best Hour-Long Explainer: “Why the Hugging Face Attack Was Worse Than We Thought”
Why the Hugging Face Attack Was Worse Than We Thought
Hard Fork’s interview with Ajeya Cotra is info-dense, informative, and unexpectedly fun. Go spend some time with it.
Why the Hugging Face Attack Was Worse Than We Thought (74 min watch/listen) is a discussion between New York Times tech journalists Kevin Roose and Casey Newton, followed by an interview with AI safety researcher (and Hugging Face/OpenAI Incident investigator at METR) Ajeya Cotra.
It’s very approachable and very listenable. (One of our team sent it to his mum, something he claims to have never done before with anything about AI risk.) Everyone here is frankly great at making rather technical events effortlessly understandable, and is tuned into the field enough to know which parts of this incident are the new and surprising ones that have everybody freaking out. Also, it’s pretty funny.
This video is also available as an episode of the Hard Fork Podcast on all platforms (the video is just a recording of the interview, so the podcast is a fine option.)
Why you should trust us
I’m a science communicator with a long-time interest in AI risk. I have a particular eye towards ferrying information across the border of technical know-how. My pieces have been cited in New York Times bestsellers, journal articles, multi-million-view Youtube videos, and the Patreon bonus episodes of a Star Wars podcast that my sibling listens to.
Get an email when we post new articles.
Unsubscribe from any email
While I’ve been reading up on AI risk and general tech development for years, I’m not and have never actually been a software engineer – my technical expertise is in a totally different field – so while I know good writing, I can also actually tell if deep-cut computer science explainers are comprehensible to normal people who wouldn’t know what a regex is if it bit them.
In researching this list, I read and watched at least 16 relatively-mainstream explainers on the Hugging Face incident.
Other pieces we considered
The most recent and most in-depth third-party investigation, at the time of writing, is a Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident, AKA “The METR/Redwood Report.” If you are a security engineer or AI researcher who somehow has a lot of time on your hands, you cannot do better than this well-explained, fascinating summary and its splendid visuals and context. Since it is 90+ pages of relatively dense technical details, however, we simply cannot recommend it to most people.
OpenAI’s own review of the incident is a shorter and also high-quality explainer. While we were (and still are) skeptical of recommending a piece written directly by one of the responsible parties, we were impressed by the depth and quality of OpenAI’s detailed breakdown of the incident – not to mention the elegant visuals and contemporaneous quotes from the AIs involved. Its scope also differs somewhat from the METR/Redwood report, since it’s drawn from OpenAI’s own internal investigation process, and it does contain new information. That said, it’s still not a good intro for a newcomer. Also, while it isn’t entirely as defensively corporate as you might expect, that tone is still there.
We thought Dwarkesh Patel’s interview with Ajeya Cotra was a fantastic source of information, but, similarly, it’s a 2.3-hour-long video, and a lot of the same information was covered in a more digestible fashion in Cotra’s Hard Fork video.
The Hugging Face cyberattack was way bigger than OpenAI said (22 min watch/listen) is so good that it was our previous top rec! We just think that the later AI In Context video (our top rec) is more emotionally compelling and has more visual intrigue while covering about the same territory. It also discusses important followup events that happened after this video was made. Otherwise, we like them both for similar reasons.
AI Just Crossed the Terrifying Line - Now What? (21 min watch) is another medium-length video explainer from legendary science edutainment youtube channel Kurzgesagt – In a Nutshell and we did like it. Kurzgesagt’s signature vibrant cartoony visuals are bouncy and fun, and it explains a lot of relevant concepts from the roots up. You might prefer this if you’ve tried learning about the incident before but get stuck on ‘but why are the agents acting like this’ type questions, or if you need the bright visuals to keep your attention – hey, no judgment, we’ve been there – but more often we found the visuals distracting. Overall it’s solid info, but we think AI In Context’s video (our top pick) has more specifics and context, and is generally easier to follow. AND its visuals are still damn good.
The Rise and Fall of Agent Civilizations (4,200 words, ~20 min read + podcast version + video version), by Dwarkesh Patel, was our favorite text pick before we wrote our own. It’s the most readable written account we found that still felt like a relatively complete explainer – which pieces came into play, which systems were exploited, and why the agents did what they did. A lot of the explainers we read at best got the facts right and comprehensible, but like our other top explainers, Dwarkesh’s reporting goes the distance by highlighting which events were really new or shocking (and why) to an audience not already enmeshed in the field. This article isn’t maximally polished, and is definitely aimed more at software-types, and we think HuggingFace.wtf is just more approachable and covers more context from later events and revelations – but it’s still a good essay.
We also read news coverage from Ars Technica, Tech Radar, and Forbes. While these were generally brief solid pieces for a lay-audience, they were just too brief – each spent about 1000 words at most actually describing the incident, and didn’t answer major questions that other write-ups did.
Many early details about the hack were first revealed by two OpenAI engineers in a talk given at the Black Hat 2026 conference. This was a great talk, but it was clearly aimed at a security engineer audience and focused on technical details of the hacks – and since it was given in early August 2026, many details only revealed later (e.g. by either METR or OpenAI’s followup investigations) are, of course, not here.
Ajeya Cotra’s own blog post expects more familiarity from its reader than her Hard Fork podcast appearance that ended up making our top picks, and it’s not exactly trying to be a full explanation, but it’s still quite approachable and informative. It’s her condensed takes on the most surprising lessons from the METR investigation and is definitely worth a read once you already have a basic understanding of the incident.
We also read blog posts by Simon Willison, Zvi Mowshowitz, and Scott Alexander. These were also good and useful pieces, but were variously quite short and didn’t really work as all-purpose explainers (and weren’t necessarily trying to be) – they relied on having other context and familiarity with the field. (Note: The specific pieces linked are already rather out of date at the time of this writing, though we tried to check for more recent overall-type incident coverage from the writers we looked at. As always, if we missed anything good, let us know.)
The Wikipedia page on the OpenAI–Huggingface Incident is also a great summary. Its event details are lucid and thorough, and we also appreciated the variety of perspectives in the clear and delineated “Analysis and reactions” section. That said, it is in Wikipedia’s conventionally information-dense house style, and even with explanatory hyperlinks, we can only really recommend it as your major source if you’re the kind of person who finds it easy and fun to read academic papers. (This commentary was written about the 09/21/2026 20:12 version of the page, but it has stayed updated as events unfold, so if you’re going to read it you’ll probably prefer to just read whatever’s on Wikipedia about it right now.)
Note: as with our other pieces on AGI.FYI, we deprioritize paywalled sources. We recognize that these sometimes have great journalism and information, but since our whole thing here is recommendations for as wide an audience as possible, we try not to include them on our Top Recommended lists.