Prompt injection. Solved.
But accidentally breaking systems is not an issue either, obviously. Even though the system prompt asks for safety rails, and other prompts wouldn't accidentally violate that.
https://www.abc.net.au/news/2026-08-10/ai-assistant-hacks-gy...
Alignment of the latest models is questionable, yes. That's a different topic.
For this particular gym incident, supposedly Opus 4.6 was used in OpenClaw, predating the current safety guardrails of Fable and co.