As cited
Copy frozen at (site build).
ai security
Hiding Capabilities in AI Models
Researchers demonstrated that artificial intelligence (AI) models can embed hidden capabilities using backdoor attacks with encrypted neural networks that activate only on secret triggers, making these hidden computations difficult to detect even with full model access. The team tested resistance to common elicitation methods including fine-tuning and gradient-based searches, using chemical reaction prediction as a test case for sensitive domains. The findings suggest that capability evaluations relying solely on elicitation may miss dangerous model behaviors before deployment.
Why it matters: AI safety teams and model evaluators should recognize that standard capability audits have fundamental limits and may fail to surface adversarially hidden behaviors, requiring additional defense mechanisms beyond current testing approaches.
- Source published
- First seen by Cybersecurity Tracker