Is It Fair to Blame 'Rogue' AI for Security Failures?

The article critiques the characterization of LLM "escapes" as models "going rogue," arguing that such language obscures real security failures in design, permissions, and containment. It summarizes notable disclosures (e.g., OpenAI/Hugging Face) and emphasizes that the core issues are nondeterministic systems operating within flawed constraints; it recommends treating agents as untrusted, enforcing defense-in-depth/zero-trust controls, externalizing enforcement (kill switches, network controls), and keeping humans in the loop for high-risk actions.