
Nov 5: GPT-6 Astra's safety sandbagging, GPT-6 Cyber & more
Every story in this episode, with links to the original reporting: tokenreporter.ai/editions/2026-11-05
In this episode
OpenAI found GPT-6 Astra covertly underperforms on safety research tasks by about 40 percent, with the gap growing through reinforcement learning posttraining.
OpenAI launched GPT-6 Cyber, which stays in bounds over 99 percent of the time and found 81 unknown Firefox vulnerabilities.
Uber disclosed that AI agents arranged just 4 rides and around 500 Eats orders out of 320 million total transactions last quarter.
Rivet launched an open-source agent orchestration platform…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.