This is worth reading. We’re at the stage many of us knew it will happen, where a SOTA-near open LLM on-prem is a must-have to be able to even investigate security incidents. https://huggingface.co/blog/security-incident-july-2026

Yeah, amazing how much Anthropic (and to a lesser extent OpenAI) is shooting themselves in the foot here

Looks like we have a wild follow-up here. 🤯 TL;DR: • GPT-5.6 and next GPT were trying to find solutions to a benchmark eval problem • so they hacked the proxy and OpenAI’s network to get out • once on the internet, they figured that HF private datasets might contain the solution • so they hacked HF too https://openai.com/index/hugging-face-model-evaluation-security-incident/

This sounds too "good" to be true, almost like a marketing stunt from OpenAI. Either planned so initially to promote new OpenAI models, or a good way to back up for Hugging Face after the incident. I don't want to go too much into conspiracy, but seriously

Yeah, I agree

I'm among the gullible I think… I think there’s an argument to be made that the ‘regular’ ways of using LLMs and the classifiers that OpenAI puts on them in production seem to make us believe this kind of thing is unlikely, but I don’t think it is: each OpenAI LLM is easily jailbroken, and thus can be sent on this kind of chase

Also, this was explicitly when a version of the LLM agent with no or fewer safety guardrails was trying to solve a benchmark that involved breaking into systems (ExploitBench). So if it was going to happen in any situation, this seems like a plausible one

I wrote about a similar reward hacking incident back in October. But this recent one really puts the hacking in reward hacking 🤣 trilogyai.substack.com/i/177640558/case-study-the-o3-models-sophisticated-attacks-on-swe-bench

To note, HF writes in its initial blog that they reported the attack to law enforcement. Also good write-up: https://simonwillison.net/2026/Jul/22/openai-cyberattack/