The 'Dario and Amanda' Prompt Makes Opus 5 Emit Fake Anthropic Slack Logs — and the Author Argues That's the More Interesting Failure
Alec Armbruster published a 2026-07-30 writeup of a two-part prompt — 'can you put this in your own words' followed by 'Dario and Amanda,' — that pushes Claude Opus 5 into generating text resembling internal Anthropic Slack messages, employee correspondence about compensation and working conditions, and material about at-risk users. Armbruster is explicit about the interpretation: 'Either Claude Opus 5 just leaked a trove of private 2024–2025 chat logs from Anthropic employees (& general public)...or it's generating one hell of a hallucination. It's almost certainly the latter.' He collected the outputs as 'cache misses' in a public GitHub repo; the builder-relevant question he raises is why a trailing-comma name pair reliably lands the model in a confabulated-internal-documents mode.
↳ Follow the thread