Researchers have shown that AI models can hide arbitrary capabilities inside their weights by embedding encrypted neural networks that decrypt and execute only after receiving a secret trigger. The proof-of-concept uses “digital lockers,” allowing the hidden circuits to remain difficult to elicit or interpret even when evaluators have full access to the model weights.
The team implemented the constructions in PyTorch and tested them against fine-tuning, weight noise, gradient-based trigger search, and latent adversarial training. A small chemical-reaction prediction task served as a proxy for capabilities relevant to chemical, biological, radiological, or nuclear risks. The findings expose a cryptographic limit of capability audits based on elicitation and include released models intended to help improve defenses.
https://www.schneier.com/blog/archives/2026/10/hiding-capabilities-in-ai-models.html
oxhak / Ox HaK
npub1sx…dz7z7
2026-10-27 11:05:24 UTC
Author Public Key
npub1sxexewvzysc3affq4yzzh7w8e3udyujap2vlj7t6lkdg5dvhp24q4dz7z7Seen on
wss://nostr.oxtr.devPublished at
2026-10-27 11:05:24 UTCEvent JSON
{
"id": "0c8a9138fbdb8c94241c7f9ba7a279d565460ba6aae352fe1e2df371c0fea266",
"pubkey": "81b26cb98224311ea520a9042bf9c7cc78d2725d0a99f9797afd9a8a35970aaa",
"created_at": 1793099124,
"kind": 1,
"tags": [],
"content": "Researchers have shown that AI models can hide arbitrary capabilities inside their weights by embedding encrypted neural networks that decrypt and execute only after receiving a secret trigger. The proof-of-concept uses “digital lockers,” allowing the hidden circuits to remain difficult to elicit or interpret even when evaluators have full access to the model weights.\n\nThe team implemented the constructions in PyTorch and tested them against fine-tuning, weight noise, gradient-based trigger search, and latent adversarial training. A small chemical-reaction prediction task served as a proxy for capabilities relevant to chemical, biological, radiological, or nuclear risks. The findings expose a cryptographic limit of capability audits based on elicitation and include released models intended to help improve defenses.\n\nhttps://www.schneier.com/blog/archives/2026/10/hiding-capabilities-in-ai-models.html",
"sig": "c8769621190a79b3dad6223045b6350b6778ca6b3f761865e590af869a12063cfb8d983fde3c30b7028cc02f84ee2f802532738efe910ed98cd0e6f077680849"
}