Partly true: Hugging Face disabled one abliterated GLM-5.3 upload and PewDiePie did say OpenAI banned him twice, but GLM-5.3 is not 'banned on every account', Anthropic's post does not call for bans, and the 'comment CRACK' hook pushes a policy-violating 753B model

The claim'They are now banning GLM 5.3 on Hugging Face on every single account.' PewDiePie 'got banned from OpenAI by trying to use OpenAI models to train his own' local model (Ajax) and 'spent weeks' on it. 'Anthropic files IPO and then immediately dropped a blog post trying to ban obliterated models,' claiming local LLMs will increase cyber attacks, and 'they're worried about GLM 5.3.' Caption adds that Anthropic found GLM-5.3 'approached Claude Mythos Preview's exploit-development results.' 'GLM got banned so make sure you get your hands on it. Comment CRACK and I'll send it over. Pirate Face has all cracked models.'

Three real news items, each bent. First, Hugging Face did disable one listing: audnai/penclaw-GLM-5.3-abliterated-for-offensive-cyber now shows 'Access to this model has been disabled. This content infringes our Content Policy.' But the same uploader's base abliterated version is still up (gated, 268 likes, 1,238 downloads last month), and the official zai-org/GLM-5.3 is live with 1.4 million downloads in the past month. 'Banned on every single account' is false; one offensive-cyber variant was pulled. Second, PewDiePie's story is real as reported: on October 2 Tom's Hardware and others covered his 'Ajax' local model and his statement that OpenAI banned him twice over distillation. That is his account; OpenAI's terms do prohibit using outputs to train competing models. Third, Anthropic's IPO prospectus was reported as leaked on September 29, the same day its Frontier Red Team published 'GLM-5.3 and the spread of advanced cyber capabilities,' so the timing jab is fair. But the post does not 'try to ban' anything. It reports GLM-5.3 at 12% on ExploitBench versus Claude Mythos Preview at 14%, says a 'meaningful threshold has clearly been crossed,' and asks governments to safety-test capable models and developers to add safeguards, while arguing defenders should get broader access. The caption gets this right; the voiceover does not. The sell is 'comment CRACK' for a DM link to a 753-billion-parameter abliterated model that almost nobody can run at home and that one platform already removed for policy violations. 'Cracked' is the wrong word too: GLM-5.3 is open-weight under Z.ai's license; abliteration edits the weights to strip refusals.

The three underlying events (one Hugging Face takedown, PewDiePie's OpenAI bans, Anthropic's GLM-5.3 cyber post the day its IPO prospectus leaked) are real and dated within the past week. The reel inflates one takedown into a platform-wide ban and a safeguards post into a ban campaign, then sells a DM link to the removed model.

Confidence

High

Posted by

an AI-news reels creator, ~6k followers; reel sent to Buddy Fri Oct 2, 2026, ~9:11 PM

See the original claim →