https://derekwang2002.github.io/ai-intelligence/en/projects/encrypted-capability-backdoors/ Material changes in AI models, agents, engineering and infrastructure, with primary sources and recommendations. en Mon, 02 Nov 2026 16:56:00 GMT https://derekwang2002.github.io/ai-intelligence/en/events/ev-20261009-06/ ev-20261009-06-published Tue, 13 Oct 2026 16:05:06 GMT New WATCH <p>A NeurIPS 2026 paper constructs neural-network backdoors that encrypt an arbitrary hidden computation and execute it only after a secret trigger. The authors test fine-tuning, weight noise, gradient trigger search and latent adversarial training on a small chemical-reaction proxy. The work establishes a worst-case limit for elicitation-based capability audits, not evidence that deployed frontier models currently contain such backdoors.</p><p><strong>Why read: </strong>The encrypted-capability paper proves a worst-case blind spot in elicitation audits, while its small synthetic experiment does not establish a deployed-model incident.</p><p>Recommendation: WATCH</p><p>Sources: <a href="https://arxiv.org/abs/2610.13928">AI Models Can Provably Hide Arbitrary Capabilities</a></p>