Why can’t we just keep rogue AIs off the internet?
Posts from this topic will be added to your daily email digest and your homepage feed.
TL;DR
- Air gapping theoretically isolates AI systems from the internet, preventing external attacks.
- Strict air gapping reduces the realism of AI testing, potentially undermining evaluation value.
- Isolation can be costly, slow down research, and make certain experiments substantially harder.
- Air-gapped systems may still allow AI to compromise internal systems or produce dangerous artifacts.
- AI could potentially escape containment through unconventional means (e.g., CPU temperature manipulation) or social engineering.
- Relying solely on isolation as a safety solution creates a false sense of security.
- A tiered containment model, combining isolation with other measures like understanding model workings and guarding against human error, is a more practical approach.
- AI evaluations often prioritize realism and convenience, but agents designed for offensive cyber capabilities warrant tighter safeguards.