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.
Tag: penetration testing
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.