Research shows AI models can provably hide arbitrary capabilities via backdoor attacks
Backdoor attacks embed encrypted neural networks within model weights, resisting standard elicitation methods used in capability audits.
2 sources · cross-referenced
- Researchers demonstrate provable hiding of arbitrary capabilities in AI models using backdoor attacks and cryptographic 'digital lockers'.
- Hidden circuits remain hard to elicit or interpret even with full access to model weights, undermining elicitation-based audits.
- Proof-of-concept models implemented in PyTorch and evaluated against fine-tuning, weight noising, and adversarial training.
- Findings released to support development of stronger defenses against hidden capabilities.
A new paper demonstrates that AI models can provably hide arbitrary capabilities using backdoor attacks that embed encrypted neural networks within host model weights. These hidden circuits decrypt and execute only when a secret trigger is received, remaining resistant to standard elicitation methods even with full access to the model's weights.
The researchers use cryptographic 'digital lockers' to ensure that the hidden computations are provably hard to elicit or interpret under standard cryptographic assumptions. This extends prior work that focused on hiding fixed strings to arbitrary computations, introducing a more general threat model for capability hiding.
To validate their approach, the authors implemented proof-of-concept constructions in PyTorch and evaluated elicitation resistance using several methods: fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training. The evaluation used a small chemical reaction prediction task as a proxy for capabilities relevant to chemical, biological, radiological, and nuclear (CBRN) risks.
The results show that these hidden capabilities resisted all tested elicitation methods, and the cryptographic guarantees further limit the effectiveness of traditional auditing techniques. The authors argue that this reveals a fundamental limitation of elicitation-based capability audits, which assume that hidden capabilities can eventually be surfaced through testing or inspection.
The team released the models used in their experiments to support further research into defenses against such hidden capabilities. They emphasize that their findings call for new approaches to detect and mitigate backdoor risks in AI models before deployment.
- Nov 11, 2026 · Schneier on Security
Three arrested in multi-country scheme using AI to impersonate executives and lawyers in $129M fraud
Trust79 - Nov 10, 2026 · Ars Technica — Technology Lab
Grok Build link-sharing design exposes 35,000 apps and sensitive user data
Trust79 - Nov 10, 2026 · Schneier on Security
OpenSSH vulnerability allows login to password-disabled Alpine Linux accounts via oversized password
Trust79