AGI Dreams Podcast – May 19, 2026

AGI Dreams – Open, Uncensored, & Local - AI News Digest · May 19, 2026 · 15m · transcript

AI Safety Internals: The Mechanics of Harm

In this episode:

AI Safety Internals: The Mechanics of Harm — Researchers found harmful content generation in LLMs depends on just 0.0005% of parameters, forming a compact cross-domain mechanism removable by pruning without degrading benign capabilities. Automated adversarial red-teaming reached 92% defense rates, while Linux kernel maintainers established standards for handling AI-generated vulnerability report noise.

Hardware Under Attack: Firmware and Embedded Security — Synacktiv bypassed Tesla Wall Connector firmware anti-downgrade…