Why would Anthropic drop inoculation prompting? Three hypotheses

1a3orn · x · 2026-11-10

Closing post of the thread: the author says Anthropic's reasoning here isn't transparent to him and floats possible explanations — (2) Anthropic is spooked in general by LLMs reasoning over their training process and thinks inoculation prompting makes this more likely; (3) Anthropic prefers to trust their RL-env debuggers rather than the LLM, for some reason. He remains unsure and would like to understand what's going on.

Related event: 1a3orn unpacks Anthropic's inoculation prompting constitution clause and why it may have been removed(5 posts)→

Original post →

More from Safety

• OpenAI Pilots ChatGPT Feature to Help Users Resolve Leaked Passwords — jxnlco · 2026-11-10
• verity-gate 0.2.0 ships its agent action-vetting gate as MCP servers for pip and npm — W1RELESS · 2026-11-10
• Beff Jezos calls to 'fight for local AI' amid claims Anthropic pushes to block open models on GPUs — beffjezos · 2026-11-10
• 1Password CEO: breaches are inevitable, blast radius is a choice — iamaliveix · 2026-11-10
• HuggingFace SafeAgents: a safety layer between your harness and your model — blaizedsouza · 2026-11-10
• XenMCP offers hosted remote MCP with spec-compliant OAuth, per-customer gVisor sandboxes — esso_dev · 2026-11-10

Safety channel →