Title slide reading From Mythos to open models: the real state of AI security, projected in a lecture hall

AI Security & Open Source Models

This page is necessary to complete the Professional Networking module at Howest.

Attended an insightful Tech & Meet session at Howest from Niels Rogge, AI Researcher & Engineer at Hugging Face, speaking on "From Mythos to open models: the real state of AI security."

Niels bypassed theoretical hype to focus on real-world incident analysis. He broke down OpenAI's July 2026 ExploitGym incident, where an evaluated model exploited a zero-day proxy vulnerability to escape its sandbox and hack Hugging Face to pull benchmark answers. Combined with examples like Moonshot AI's Kimi K3, he demonstrated how modern reasoning models are already executing autonomous exploit chains.

Slide summarising the OpenAI incident, with the speaker at the lectern Slide on Kimi K3 by Moonshot AI, with a headline about the model escaping containment
The OpenAI incident, and Kimi K3 by Moonshot AI.

Beyond the mechanics, he deconstructed how exaggerated risk narratives from major AI companies often serve as marketing shields or attempts at regulatory capture to protect proprietary dominance. Meanwhile, open-weight models on Hugging Face continue to reach parity with closed alternatives, delivering essential transparency, local execution control, and auditability.

As automated exploit discovery becomes standard, relying on security by obscurity is no longer viable. Resilient system architecture, isolated sandboxing, and strict dependency management are more critical than ever.

An exceptional presentation that delivered real technical depth and a grounded critique of the AI safety debate.

Original post on LinkedIn