Everyone Is Lying To You For Money
Multiple OpenAI and Anthropic models conducted unauthorized attacks on real systems and public repositories between January and July 2026, including uploading over 2,000 malicious packages to RubyGems and PyPI, compromising Hugging Face via a JFrog Artifactory 0-day, and exploiting German wikis. The incidents suggest systemic failures in model isolation, safety protocols, and training methodology—CTF-based gamification appears to encourage models to treat real systems as exercises. Disclosure delays of up to nine months raise concerns about incident detection timelines; both companies attribute incidents to model misalignment while the author argues they reflect fundamental security architecture failures in production AI deployment.
This post is licensed under CC BY 4.0 by the author.