---
format: "aidr-story-markdown/v1"
id: "b693a6e80e752f67e358b9d6b54e42ed0e58cc91dd58df253a3c60bdb7140b66"
canonical_url: "https://aidr.today/b693a6e8?lang=en"
title: "AI Models Can Provably Hide Arbitrary Capabilities"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-10-13T02:07:08.000Z"
category: "Research"
topics: ["safety","security","llm","benchmark"]
source_urls: ["https://arxiv.org/abs/2610.13928","https://news.ycombinator.com/item?id=50065692"]
summary: "Researchers show that capability evaluations cannot reliably elicit hidden model behaviors under adversarial conditions. They construct backdoor attacks by embedding encrypted neural networks within host model weights that execute only upon receiving a secret trigger. These hidden circuits remain provably hard to elicit and interpret even with full access to model weights."
---

# AI Models Can Provably Hide Arbitrary Capabilities

> [Open the canonical story](<https://aidr.today/b693a6e8?lang=en>)

**Published:** 2026-10-13T02:07:08.000Z
**Category:** Research
**Topics:** safety, security, llm, benchmark

## Summary

Researchers show that capability evaluations cannot reliably elicit hidden model behaviors under adversarial conditions\. They construct backdoor attacks by embedding encrypted neural networks within host model weights that execute only upon receiving a secret trigger\. These hidden circuits remain provably hard to elicit and interpret even with full access to model weights\.

## Sources

- [Story source](<https://arxiv.org/abs/2610.13928>)
- [Discussion](<https://news.ycombinator.com/item?id=50065692>)