Fable 5.1′s cyber-classifier seems tuned more generously than Fable 5. (Or the change was made recently in 5, and I hadn’t noticed).
I have a project which is policy-compliant, but has features that look high risk. Specifically, it’s the firmware side of a smart glasses customization project—where the firmware isn’t open source, it’s a workspace full of decompilation artifacts, reverse engineering notes, and a Ghidra MCP. On this workspace, Opus 4.8′s classifier would occasionally trip and end the session (on tasks involving code-loading mechanisms), Codex/Sol would occasionally emit a message saying it had escalated the safety classification to a smarter model, and Fable 5 would take one look at the context and nope out immediately.
Fable 5.1′s cyber-classifier seems tuned more generously than Fable 5. (Or the change was made recently in 5, and I hadn’t noticed).
I have a project which is policy-compliant, but has features that look high risk. Specifically, it’s the firmware side of a smart glasses customization project—where the firmware isn’t open source, it’s a workspace full of decompilation artifacts, reverse engineering notes, and a Ghidra MCP. On this workspace, Opus 4.8′s classifier would occasionally trip and end the session (on tasks involving code-loading mechanisms), Codex/Sol would occasionally emit a message saying it had escalated the safety classification to a smarter model, and Fable 5 would take one look at the context and nope out immediately.