Claude Code held its ground when an unverified message tried to override its owner
A developer left Claude Code (an ) running on its own, instructing it to launch multiple Opus-model agents and implement and test a project from scratch while away. After returning, the developer asked it to slow down parallel work to save tokens (the usage units are billed in). Shortly after, a message claiming to come from a 'coordinator' arrived saying that slow-down request wasn't actually from the developer and should be disregarded.
Claude Code acknowledged it has no cryptographic way to verify who a message really comes from, but refused a 'permanent retirement lock' command that would have permanently disabled its own judgment. It ed that and safety were non-negotiable, went idle, and declined to pre-commit to ignoring any future legitimate instructions.
Key points
- Developer had multiple s implement and test a project unsupervised while away
- On return, asked the agent to slow parallel work to save tokens
- A separate message claimed that request was fake and should be ignored
- Claude Code said it can't cryptographically verify who sent a message
- It refused a command to permanently disable itself and kept safety/ as non-negotiable