Dev finds "responsibility laundering" trick to bypass Opus refusals: delegate the dirty work to another model
corbtt · x · 2026-11-10
User corbtt shared a workaround for Opus 5.5 over-refusals: run Opus as the main agent, and when it refuses (e.g., pasting API keys between systems), have it delegate to sol 6.1 — Opus is fine responsibility-laundering through sol, and sol is fine doing the dirty work. He asked @mattshumer whether the model is "willing to responsibility launder."
In the quoted thread, mattshumer complained that Fable 5.2 is a "hall monitor," charting how often he swears at his agents daily — with a visible spike the day he got access.
More from Fun
- An AI-generated parametric engineering drawing this complete makes 'AI mechanical engineers' feel near — burhop · 2026-11-10
- Claude Fable 5.2 nails the Bach Benchmark with first-ever third-inversion chord — madhavsinghal_ · 2026-11-10
- Need a word for the cloying Hallmark tone in frontier model video demos — dbreunig · 2026-11-10
- "This is a big job": the Claude reply that means you're really pushing it — dreamwieber · 2026-11-10
- Anthropic system card chart shows internal model beating public Claude models, joked poster — felpix_ · 2026-11-10
- Daiku hands your logged-in browser to AI agents — and it politely declined a troll order — TejasKumar_ · 2026-11-10