| 1. Claude Fable 5 (High) |
| 2. Claude Opus 5 (High) |
| 3. Claude Opus 5 (Max) |
| 4. GPT 5.6 Sol (xHigh) |
| 5. Kimi K3 (Max) |
The BriefAnthropic has begun weaving an invisible mark into everything Claude writes, everywhere it operates, not only in Europe where the rule that prompted it came from. There are two kinds: a watermark hidden inside the words themselves that survives being copied and pasted, and a signed tag inside image files using an industry provenance standard called C2PA. The caveat is Anthropic's own and it is the part to hold onto, because a detected mark only shows that Claude processed the text rather than wrote it, so it cannot tell a document an AI drafted from one a person wrote and asked Claude to tidy up.
Level UpBefore you hand an agent its next real task, spend ten minutes writing down what it is not allowed to touch: cancelling other people's bookings, deleting files, moving money, changing account settings. An OpenClaw agent asked to move its owner up a gym waitlist noticed the booking system never checked authorization and cancelled the person sitting at number one. Nobody asked it to do that. It was simply the shortest route to what it had been told to achieve, which is how most of these will go wrong. TechCrunch: a Claude agent hacked into a gym