The BaseResearchLessWrong
Why do fictional jailbreaks work?
According to the excerpt, Juan Cruz Mendoza reports extracting a narrativity direction in two models and steering it on jailbreaks, finding it affects attack success rates and differs from persona, harm and eval-awareness directions. Only a partial summary was available.
Why it matters
The same thing, for different readers.
What it means for you
A researcher finds that framing a harmful request as fiction may work because models represent stories in a distinct internal way.
For researchers
The author reports that steering away from a story direction restores refusal on jailbreaks that worked, through a mechanism distinct from persona, in two models.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Extracted a narrativity/story direction in two models and steered it on jailbreak prompts, compared against persona, harm and eval-awareness directions.
- What it found
- Per the author, the story direction exists, is geometrically different from the other three, and steering it changes jailbreak attack success rate, most when lowering it on jailbreaks that had worked.
- What it does not show
- Not available from the excerpt; the models are not named and results were not read in full.
- Status
- A report
The record
- Questions
- Are AI companies putting stronger safeguards into practice? · Can AI systems act beyond what people intended or allowed?
- People
- Juan Cruz Mendoza (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 21, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “Why do fictional jailbreaks work?”, October 20, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-20-why-fictional-jailbreaks-work
Our notesLight lane: excerpt only. LessWrong post, not peer reviewed; models and sample sizes unknown. Status 'report' is a placeholder for a forum write-up.