Claude Code Auto Mode: A Safer Filter With an 80% Bypass Claim
Send us Fan Mail
Anthropic says Claude Code auto mode catches far more harmful actions than human reviewers, but a prompt injection researcher now claims an 80% bypass rate. Augur explains how both findings can be true, why a model-based classifier is still only a filter, and what hard access boundaries businesses should put around coding agents that read untrusted text.
The episode also examines Google's double-blind model evaluation, OpenAI's containment post-mortem and partial outside review, and an AI-assisted breast cancer research finding that still needs clinical validation…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.