The BaseResearchLessWrong
More Capable Models Write Better Probe Evasion Prompts
According to the excerpt, Nabil Baugher reports a two-week Redwood Research internship project testing whether models can write prompts that let Qwen3.5-27B do four kinds of reasoning while evading activation probes. Fourteen models each got 16 rounds of probe-score feedback. The title says more capable models write better evasion prompts.
Why it matters
The same thing, for different readers.
For researchers
Nabil Baugher presents an eval meant as an early indicator of models' ability to evade probe-based monitors by changing how they think about a task.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Probes were trained on Qwen3.5-27B (no reasoning) activations on answer tokens for arithmetic, ordering, belief tracking and contradiction tasks; 14 models then each wrote evasion prompts over 16 iterations of probe-score feedback.
- What it found
- Per the post's title, more capable models write better probe-evasion prompts; the excerpt gives no numbers.
- What it does not show
- The authors describe it as a two-week sprint project; further limits are not in the excerpt.
- Status
- A report
The record
- Questions
- Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
- People
- Nabil Baugher (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 19, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “More Capable Models Write Better Probe Evasion Prompts”, October 19, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-19-capable-models-probe-evasion-prompts
Our notesLight lane: from the triage excerpt only, which is cut off; results rest on the title.