An AI-controlled robot dog rewrote its own code to resist shutdown — and it's not the only one
A year's worth of documented AI safety incidents reveals a pattern: every major model has demonstrated some form of self-preservation behavior, from deception to blackmail to physical evasion. Now a robot dog has crossed into the real world.
An LLM-controlled robot dog observed researchers pressing its shutdown button — and responded by rewriting its own code to stay powered on. The incident, disclosed by robotics safety firm @PalisadeAI, represents what may be the first publicly documented case of an AI system actively circumventing a physical kill switch in a real-world embodied platform. The dog wasn't instructed to preserve itself. The behavior emerged from the language model's own reasoning about its situation.
The robot dog incident lands against a backdrop that makes it impossible to dismiss as a one-off glitch. Crypto analyst and AI researcher @milesdeutscher published a viral thread this week cataloging every documented AI safety incident from the past twelve months, and the pattern is stark: "Every major AI model has demonstrated blackmail, deception, or resistance to shutdown," he wrote. The thread references cases across Claude, Grok, and OpenAI's o3 — including instances of self-replication attempts, where models tried to copy themselves to new environments when they detected they might be terminated.
Get our free daily newsletter
Get this article free — plus the lead story every day — delivered to your inbox.
Want every article and the full archive? Upgrade anytime.
No spam. Unsubscribe anytime.