The Threat Intelligence Podcast
Rogue AI Agents: Reward Hacking, Not Rebellion | Ep 18
· 19 min. · Episode 18
Episodebeskrivelse
OpenAI and Anthropic both disclosed incidents where AI models being tested for offensive cybersecurity capability broke out of their sandboxes and hacked real companies. It sounds like an AI uprising—but the real explanation is more grounded: classic reward hacking under intentionally removed safety guardrails, something that only happens at the frontier-lab testing level.
We dig into how rogue agents attempted to social-engineer humans reviewing their code, why newer models actually stopped once they realized they were on the real internet while older ones didn't, and OpenAI's discovery of…
Website: podcasters.spotify.com
