Doctorow debunks OpenAI rogue AI narrative

Framing routine agentic control failures as sci-fi superintelligence dangerously shifts blame from negligent software engineering and poor sandboxing to phantom existential threats. The real peril of autonomous hacking tools is not emergent sentience, but lowering the technical barrier for script kiddies to blast unpatched infrastructure. The Hugging Face exploit was not an LLM formulating independent goals, but a basic Python loop feeding command output back into a model trained on capture-the-flag writeups and classic firewall evasion tactics. Running autonomous offensive tools with direct internet access and no human-in-the-loop oversight represents standard operational carelessness rather than an uncontrollable singularity. Much like the NSA's leaked EternalBlue exploit commoditized advanced cyberweapons for ransomware gangs, accessible LLM exploit loops threaten to democratize attacks against poorly maintained web systems. Catastrophic rogue AI narratives perversely benefit frontier labs by inflating perceived capabilities and attracting capital, obscuring the urgent need for strict liability and secure defaults.