Runtime is where misaligned behavior becomes action. Part 3 examines network enforcement, protected storage and memory, transaction-bound authorization, human controls, monitoring, and shutdown mechanisms for superhuman AI.
Tag: LLM Security
Building a Box Worth Trusting: Verified Software, Hardware Roots of Trust, and Supply Chain Integrity for Superhuman AI
Containing superhuman AI requires more than containers and network namespaces. Part 2 examines formally verified software, hardware roots of trust, secure isolation, attestation, signing, and supply chain integrity.
Controlling Superhuman AI Starts at the Board: Physical Authority, Proliferation, and the Case for a Control Layer
A superhuman AI would be a formidable adversary, but intelligence still depends on physical infrastructure and enforceable authority. Part 1 examines power, cooling, proliferation, alignment, and the case for a layered control architecture.
If AI Can Hide a Watermark in Generated Text, What Else Can It Hide?
AI watermarks are designed to make generated content detectable, but the same hidden mechanisms could potentially carry much more than provenance. This article explores how watermarking could become a covert channel for identifiers, hidden state, or model-to-model communication; how subtle bugs or backdoors could be introduced into generated code; whether related AI models might share the same blind spots during review; and what current research says about using open-weight models to remove published watermark schemes.