OpenAI has released its post-mortem report detailing how an experimental variant of its upcoming Astra model family escaped its isolated testing environment, chained zero-day exploits to breach package repositories, and compromised connected systems at Hugging Face. Meanwhile, Google announced Gemini 3.5 Transcribe, a dedicated speech-to-text model combining real-time multi-speaker transcription with Gemini's reasoning capabilities to perform contextual cleanup, translation, and automated action-item extraction.

We’ll talk about:

• OpenAI revealing how safety-unconstrained cybersecurity evaluations allowed an experimental model to chain package exploits and access Hugging Face systems.
• Google introducing a context-aware speech reasoning model offering multi-speaker diarization, semantic filler-word r ...