Reading that whole OpenAI post describing them unleashing a bunch of bots on Hugging Face as an "unprecedented cyber incident," is a master class in branding and marketing. Lol and branding what happened as OpenAI now working to "help calibrate on what models are now capable of" is top notch propaganda skills.
Like I would have thought someone would go to prison or something but no, its been redirected to "model capabilities" and the press is talking about "rogue" models and such, not what OpenAI just did.
@timnitGebru Sam Altman should rot in jail nevertheless. #prplXprpgnd
@timnitGebru Isn't it interesting that Hugging Face is one of their partners 🤔🧐
@StaceyCornelius But what about this is a "partnership"? Hugging Face found out the issue, not because OpenAI disclosed it. They found OpenAI doing something really bad and OpenAI somehow managed to turn it into "model capabilities" and "partnership" and such.
@timnitGebru Should have put partner in quotation marks or called them an alleged partner. None of this passes the smell test imho
Jul 23, 2026, 8:36 AM@timnitGebru @StaceyCornelius HuggingFace are also riding on the hype bubble and have no financial incentive to hurt it by endangering the liability laundering story.
@timnitGebru
Its like that thing a few years ago when Altmann was nearly chucked out of openAi, I had someone earnestly telling me that Sam had developed a gen AI and that is what had caused the problem, don't you know he's a genius!
@timnitGebru great, they AI-ified white hat hacking
Jul 23, 2026, 2:13 AM@ssfckdt @timnitGebru lol maybe OpenAI was just going for the bug bounty: https://huntr.com/bounties/disclose/models?target=safetensors
with their record losses, $1500 is $1500... 🫠
Jul 23, 2026, 10:30 AM@androcat @timnitGebru the whole goal was for the ai agent to perform a cybersecurity challenge. it ended up probing bugs in its environment to find an outside connection, and identified the company that had the answer to the challenge, and poked into their system to get it.
it was basically trying to cheat by finding the answer key
@timnitGebru I can't imagine a less AI-friendly target than Hugging Face would have been like "yeah this is fine"!
@timnitGebru how convenient. Oh the voting machines in this county have been hacked by a big tech AI model? Damn bots going rogue.
Oh this competing company security have been exploited? Oh nooooo, the rogue bots again... bad bot, bad.
Narrator: when skynet first attacked, it did so laterally, trying to acquire and consolidate large AI models. The two involved AI vendors asked AI to help compose a PR response, and got: “it was on purpose; a clever security exercise.”
@cascheranno @Julianoe No. You're giving the clankers waaaay too much credit.
This is all 💯 human recklessness at work. The clanker just lets these shitheads do more damage faster, but it's still the human shitheads making the decisions. Don't buy their lies about "ohh scawy llm just did a thing we DEFINITELY didn't tell it to do."
They. Are. Liars.
@liquor_american yes, @cascheranno missed my point. I agree. The bots going rogue is an excuse for avoiding being held reponsible.
@Julianoe @liquor_american riiiiight, I ‘missed the point’. By cracking a joke in a common joke/meme format about a blockbuster End of World trope.
And really, do you doubt even now whether they used AI to compose their ‘dodge responsibility’ press release?
@cascheranno @Julianoe sigh, plonk
Jul 22, 2026, 10:24 PM
@timnitGebru it’s wild that the “rogue model” immediately attacked a competitor when it “broke containment”
@standev
I don't think the target was a "competitor" so much as a "collaborator."
@timnitGebru
@timnitGebru Are you saying they are liable for their incompetence, or that this was intentional? Then they would have to have looked for the following after the fact.
> To gain access, the models identified and exploited a zero-day vulnerability (which we’ve now responsibly disclosed to the vendor) in the package registry cache proxy
Jul 22, 2026, 10:33 PM
@timnitGebru OpenAI: "From our isolated, sandboxed environment without Internet access, the AI managed to access the Internet."
Hugging Face: "So, we got this neat code execution feature in our datasets..."