Can AI models hide secret messages in images and read them out again later? That is exactly what researchers investigated, as t3n reports. According to the report, the AI did not invent any new methods in the process. Even so, the findings could matter for AI safety.
The article's headline indicates that two systems were involved in the experiment: OpenAI's Astra decoded an image message that Anthropic's Fable had created. The short version available to us gives no further details.
What hidden image messages are about
Hiding information in inconspicuous carriers is called steganography. Unlike classic encryption, the goal here is not only to make the content unreadable. The very existence of the message is meant to go unnoticed. An image can look completely ordinary to people and still contain additional information.
The fact that the models did not develop new techniques is an important piece of context. They reportedly relied on approaches that were already known. That argues against a sudden technological leap, but it shows that they can apply such techniques and also decipher them again.
Why this matters for AI safety
AI safety is concerned, among other things, with being able to understand what models do and how they pass on their results. If systems package information in a way that people do not notice, this can make oversight harder. That applies above all when several AI systems exchange data.
For teams that use AI tools, a practical question can arise from this: is only the visible text of an output checked, or also whatever might be hidden in images and files? This is not a blanket warning, but rather a reason to rethink one's own review processes.
Whether and how strongly this behavior shows up in everyday use cannot be derived from the short report. Anyone who wants to know more should read the full article at t3n and consult the study itself, provided it is linked there.
