On Moltbook

Bruce Schneier’s recent posts collect multiple, intersecting security risks driven by modern AI and adversarial behavior: LLMs can practically deanonymize users at scale by inferring attributes from a few posts and searching the web; training-data poisoning is trivial (a single webpage can influence major chatbots); and LLMs produce highly predictable, low-entropy passwords. Separate non‑AI stories show real-world exploitation: North Korean actors are running fake-recruiter coding challenges that install malware on developer machines, and researchers found ways server-side control or recovery/