Breaking01 / AI
Internal signals detect medical AI manipulation more often
Researchers ran the tests without retraining, using an existing tool to turn internal signals into text and a general-purpose model to evaluate it.

Researchers reported on 20 Nov 2026 that evaluating a medical AI model’s internal signals detected manipulation more often than reviewing its answers in all eight attack conditions they tested.
In the paper, the researchers described how they converted Gemma-3-27B’s internal signals into text. A separate model evaluated that text and the answers independently, without seeing the original input.
Both methods rarely flagged manipulation when the inputs were unaltered, according to the paper.
The researchers used an existing tool to turn the signals into text and a general-purpose model to assess them. The experiment required no retraining.
The paper also reported that internal signals detected manipulation more often than answers in a test with Llama-3.3-70B in which the model was induced to omit a lesion.
The abstract did not disclose detection rates for either method by test condition.