🚨 AI Safety Crisis Analysis: Behind the ">10% extinction" curtain: Liability Evasion An Anthropic researcher quit. A safety lead backed him. 100 million views. The public hears an alien mind. Safety people hear confirmation of the story they already told themselves. The press release writes itself. The actual system? That's a different story. The problem remains software. [ The play ] The number is not the story. The story is who stands in court when the software fails. The law already has a category for this. It is software. Designed, trained, deployed by a company. When it breaks things, the company is the defendant. The extinction quote does not add a duty. It moves the argument off that duty. Off product liability. Onto a story only the vendor gets to write. [ The panic ] A number dropped. >10% extinction. Headlines followed. The number is not a measurement. It is a personal estimate from someone paid to work on the problem, said out loud on the day a colleague quit over the same fear. That is newsworthy. It is not evidence. [ What the system actually is ] Inference is local. One query. One context window. One sample from frozen weights. A "swarm" is a single system making many inference runs. Each run starts from scratch. No carry-forward state. No persistent memory. Stop the process and the cast disappears. Same model. Different prompts. A harness. Role-play until the compute stops. The system never left the premises. There is no second party to blame. No mind to put in the dock. Only an operator. [ Why the narrative fails scrutiny ] Unauthorised access is not a rogue agent. It is a poorly secured perimeter. Tools in a loop. No checkpoints. Context with the same privileges as actions. That is poor systems design. Not alignment. A distribution is not an agent you control with RL. It is a mathematical object, not a mind. No self. No goals. No psychology. Put this in front of a software engineer, a safety engineer, or a cybersecurity specialist. They will ask: where is state persisted? Where is the evaluation layer? What happens if you stop the process? Where is your monitoring or escalation protocol? The answers are boring. It does not survive a panel of expert auditors. These point to negligence. [ The game ] Labs cannot certify what they commercialise. So they rebrand missing controls as "misalignment" and missing adults as "existential risk." That is the oldest trick in the book. Fear first. Then market and policy capture. A special regime. Alignment boards. Lab-written rules. Anything but the category that already applies: the manufacturer of the software. Regulators get the panic, not the architecture. Nobody in the room asks who shipped it. Lobbies make sure they only hear one side. Change the public's picture of the product and you soften the law that already makes the lab responsible. Autonomy. Personhood. Alignment. None of these is true. All of them move the defendant offstage. [ What actually works ] Stop the compute. Remove the context. The agents were never there. Secure the perimeter. Constrain tools. Verify actions outside the model. Enforce the duty that already exists. Make labs liable when the software they ship breaks things. Fines when safeguards are not enforced. Signs of negligence get treated as negligence. Regulate with subject-matter experts. Not a community that needs the model to be a mind so the vendor can stop being the operator. Sampling ≠ agency. Soft programs ≠ personhood. A frozen distribution going rogue is safety theatre. ☕️ How many more extinction quotes will treat a poorly secured system as a menacing hive before someone puts an adult in the room?