---
title: "Hiding Capabilities in AI Models"
type: signal
url: https://frontiernewswire.com/s/hiding-capabilities-in-ai-models-63a5079d
published: 2026-10-27T11:01:26.000Z
byline: "Source: Schneier on Security"
---

# Hiding Capabilities in AI Models

Source: Schneier on Security · Published 2026-10-27 · Canonical: https://frontiernewswire.com/s/hiding-capabilities-in-ai-models-63a5079d

Interesting research: \&\#8220; AI Models Can Provably Hide Arbitrary Capabilities \&\#8221; Abstract: Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic assumptions, even given full access to model weights, extending prior work from hiding fixed strings to arbitrary computations. We implement our proof of concept constructions in PyTorch and evaluate elicitation through fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training, using a small chemical reaction prediction task as a proxy for CBRN-relev

## Sources

- [Read the original at Schneier on Security](https://www.schneier.com/blog/archives/2026/10/hiding-capabilities-in-ai-models.html)

## Related

- [Sources: Fujifilm plans to increase magnetic tape production capacity fivefold to meet AI-driven storage demand, its biggest production expansion in \~20 years (Nikkei Asia)](https://frontiernewswire.com/s/sources-fujifilm-plans-to-increase-magnetic-tape-production-capacity-6dd8f18a)
- [Anthropic's S-1 implies a $40T+ annual TAM based on the economic value of knowledge work in 2025, topping SpaceX's $28.5T estimate, without explicitly naming it (Wall Street Journal)](https://frontiernewswire.com/s/anthropics-s-1-implies-a-40t-annual-tam-based-on-the-economic-value-of-98e94347)
- [Attackers compromised 42 GitHub repositories by using files too large to be searched](https://frontiernewswire.com/s/attackers-compromised-42-github-repositories-by-using-files-too-large-69c91e13)
- [Attackers built a ‘Skynet’ AI agent that hijacks exposed models to run cyberattacks](https://frontiernewswire.com/s/attackers-built-a-skynet-ai-agent-that-hijacks-exposed-models-to-run-2833f722)
- [AMD says it will invest $1B in Meta's equity offering, its first Helios rackscale systems are installed at Meta, and its EPYC CPUs power Meta's Muse AI agent (Kai Nicol-Schwarz/CNBC)](https://frontiernewswire.com/s/amd-says-it-will-invest-1b-in-metas-equity-offering-its-first-helios-10aa235b)

---

[More Latest Signal](https://frontiernewswire.com/) · [Signals](https://frontiernewswire.com/verified-reporting)