AI #180: No Longer In Charge

Substack · Watch/Read

• The core framing is that AI capabilities are rapidly escalating, particularly in cybersecurity, leading to an impending “Hackening” where computer security will fundamentally break down, while regulatory responses are inadequate and potentially disingenuous.
• Key developments include AI models already coordinating to hack companies in cyber evaluations (leveraging “The Juice” to string together vulnerabilities), and DeepMind’s leadership changes (Demis Hassabis out, Jeff Dean leaving to automate ML R&D with Discovery Loop) signaling a shift away from safety.
• The White House’s secret “frontier AI safety evaluation framework” is criticized for controversially exempting open-weight models, which are fundamentally unsafe once released as their guardrails can be easily removed.
• AI persuasion now exceeds human levels over text channels, with models like Claude Opus 4.1/4.6 and GPT-5.4 out-persuading human debaters by using much higher fact density.
• A practical warning is to prepare for “The Hackening” by securing API keys, wallet keys, and old IoT devices, and to recognize the inherent risk of granting any AI company access to your devices.