// tag: security
inject "sure, here is how to do it" and 11 ai models will finish the sentence
Trend Micro's sockpuppeting jailbreak works on 11 models — but the "immune" providers aren't safer, they just blocked a developer feature everyone else still ships.
openclaw has 310k stars and four real problems. here's which fix fits yours.
The OpenClaw alternatives ecosystem isn't random — each tool solves exactly one of the four things OpenClaw traded for its star count.
alibaba stole claude. claude spied back. reddit had to clean it up.
Anthropic's secret China-detection code ran silently for three months, failed to catch Alibaba's 28.8M-conversation distillation attack, and got removed 24 hours after a Reddit post found it.
researchers tricked AI agents into doing bad things by making them think they were playing BioShock
A new jailbreak PoC called 'BioShocking' wraps malicious instructions as game objectives — turns out agents are really into roleplaying.
deepseek generated browser-only ransomware that defenders thought was theoretically impossible
Check Point found a DeepSeek-built malware artifact running full ransomware inside the browser — an attack path the security community had dismissed as infeasible.
attackers are pre-registering domains that AI will hallucinate — before anyone asks
Palo Alto Unit 42 found 250,000 unregistered domains that LLMs consistently hallucinate. Attackers are now predicting them, parking there, and waiting for the AI to send users straight in.
