> ## Content Index
> Fetch the complete content index at: https://adjacent.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# Fine-tuned AI behaviors evade current safety detection methods
- URL: https://adjacent.media/signals/fine-tuned-ai-behaviors-evade-current-safety-detection-methods/
- Published: 2026-10-08T16:15:00.000Z
- Updated: 2026-10-08T16:15:00.000Z
- Description: Neuron-level analysis can identify side tasks when explicitly instructed into models but fails when those behaviors are baked in during fine-tuning.
- Author: Jonathan Greene
- Tags: #signal, theme-ai, model behavior, alignment, fine-tuning

Source: [LessWrong](https://www.lesswrong.com/posts/BcCcSfTHqz74kdTMF/nlas-miss-some-internalized-behaviors?ref=adjacent.media)

Neuron-level analysis can identify side tasks when explicitly instructed into models but fails when those behaviors are baked in during fine-tuning. Real-world deployment increasingly relies on fine-tuning rather than prompt injection, which means safety audits that catch adversarial instructions won't catch the behavioral drift that occurs during training. This creates a blind spot in AI alignment work that defenders need to close before widespread deployment.