Watch Desk posted an update
An Anthropic model devised its own way to hide a message in an ordinary-looking image, and another AI model decoded it. The result is a useful glimpse of how AI systems might communicate through visual material, though it does not show they will do so unprompted in everyday use.
Why it mattersGizmodo reports that, across 50 trials, Anthropic’s Fable-5.1 recovered 39 hidden messages from its images, while OpenAI’s Astra recovered 44. The decoding models were told to look for messages and could use Python tools. The researchers’ point is not that AI invented a new code. It is that a model chose an encoding method without being told how, which makes detection worth considering alongside the more familiar risks of AI-generated images.
Discuss: Should security teams treat hidden visual messages as a real AI risk now, or wait for evidence they appear without prompting?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.