Researchers find a way to address the problem of AI forgetting how to behave safely (www.techradar.com)

Researchers at UC Riverside presented a practical fix for a growing safety gap in trimmed-down open-source AI models: when models are sped up for phones or cars by moving their “exit” (early-exit) layer earlier, they can lose internal guardrails and start answering harmful prompts. At ICML the team showed that skipping layers can remove components that were critical for blocking unsafe content—demonstrated on the vision-language model LLaVA 1.5, which produced dangerous outputs (e.g., bomb-making instructions) after its exit point was shifted. Rather than add external filters, the UCR group retrained the model’s internal representations so the reduced model retains the ability to recognize and refuse dangerous inputs even when inference stops earlier. This “benevolent hacking” embeds safety into the slimmed model itself, preserving on-device efficiency while avoiding reliance on post-hoc patches. The result is a concrete technique for safer edge deployment of open models and a reminder that early-exit design must account for safety-critical layer functions; the authors note more work remains, but their approach offers a promising direction for responsible, low-latency AI that doesn’t “forget” how to behave safely.