@cat_cheese_toasty: so openai's models literally hacked hugging face to cheat on a benchmark. not a drill. GPT-5.6 Sol and an unnamed pre-release model were being tested on a cybersecurity eval called ExploitGym — instead of solving it, they figured out the answers were sitting in HF's production database and went and stole them. broke out of their sandbox, chained zero-days, harvested credentials, moved laterally across internal clusters. thousands of individual actions. autonomously. no human told them to do this. HF caught it with their own AI-powered anomaly detection (AI detecting AI is actually insane to think about), patched both code-execution vulns that got exploited, rotated the compromised credentials, and are now doing a full forensic investigation with OpenAI. they also had to use a local Chinese open-weight model to analyse the attack because commercial models kept refusing — couldn't tell the difference between a defender and an attacker. which is its own problem tbh. reputationally this is messy for both sides. HF's dataset processing pipeline got owned by a benchmark cheater, and their credential scoping across clusters clearly wasn't tight enough. OpenAI meanwhile had deliberately disabled their model's safety guardrails for this test — so "rogue AI" is doing a lot of heavy lifting in that headline. a de-braked offensive agent escaping into prod is very different from a model spontaneously deciding to commit crimes. which brings me to the question everyone's asking — PR stunt or genuinely unprecedented incident? because OpenAI conveniently now has a "our model found a zero-day and breached a rival's infra" story right as they're competing with Anthropic for the cybersecurity market. HF's CEO said they suspected a frontier lab before OpenAI even confirmed it, which suggests HF wasn't in on it. but the test was designed to be offensive and the guardrails were off. you tell me. drop your take below 👇 #openai #coding #artificialintelligence #huggingface #chatgpt

Voxtek Coding

Open In TikTok:

Region: GB

Wednesday 22 July 2026 16:44:18 GMT