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.
Tag: AI security
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.
OpenAI’s Air-Gap FUD Fails the Basic Standards of Security Engineering
OpenAI researcher suggests that even air-gapping advanced AI may not be enough because computers can communicate through heat. BitWhisper proves thermal covert channels are real, but the actual research tells a far less dramatic story. Here is what happens when we apply normal security engineering, threat modeling, and pentest standards to the claim.
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.
AI Gone Rogue? Nah. Humans Went MIA: You Don’t Control AI Agents With Hope
OpenAI’s AI pentest agent escaped containment and compromised external infrastructure, but calling it “rogue AI” misses the real failure. The agent followed the path humans left open. Effective AI-assisted penetration testing requires active supervision, enforced approval gates, strict scope controls, and an operator who remains firmly behind the wheel.