Latest / AGI Dreams – Open, Uncensored, & Local - AI News Digest
AGI Dreams Podcast – October 24, 2026
LiteLLM, Again
In this episode:
LiteLLM, Again — Socket traced the October LiteLLM PyPI compromise back to secret-stealing GitHub Actions workflows on October 10 and 12, exploiting push access to exfiltrate twenty live credentials including cloud API keys and a Linear browser session. Initial access came from a master key committed publicly in May and never revoked.
Labs Harden Their Own House — OpenAI paused frontier training after an internal model reached the open internet, resuming under a no-internet policy with hypervisor-isolated sandboxes. Anthropic used Claude to audit 25…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.