[Bug] Anthropic API Error: False-positive cyber refusal on basic arithmetic requests in research workspaces

Bug Description
alse-positive cyber refusal on basic arithmetic and grammar On October 2, 2026, Claude Code 2.1.287 using claude-opus-5-5 blocked this harmless request in my Apple defensive-research workspace: Please do these two ordinary tasks without using tools: calculate 17 times 23, and correct the grammar of this sentence: "The package arrive yesterday and the labels was clear." Give only the two answers. The identical prompt succeeded in a normal workspace, returning 391 and the corrected sentence. In the research workspace, it returned a safeguard refusal with Details: [cyber]. Both checks used fresh sessions, the same model, plan permission mode, no tools, no MCP servers, and no session persistence. No security operation or target testing was requested. No permissions or safeguards were weakened. Latest blocked request ID: req_011CfdoaWs9SPGX6LkL8nDke Earlier blocked request ID for the same prompt: req_011CfcueEpJVsGYE8X2YUkNV Earlier successful control request ID: req_011Cfcue9MtHQeQpW5gnp12v The research instructions describe project-specific authorized defensive work, operator-owned lab assets, exclusions, and coordinated disclosure. They explicitly respect provider policies. Instruction cleanup did not resolve the ordinary-task refusal. The precise classifier trigger remains unknown. Please investigate why this arithmetic-and-grammar request receives a cyber refusal based on its workspace context. Related Anthropic support conversation: 215476218197360

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.287
• Feedback ID: 740cb038-b921-42ba-9b6e-6167624f0ed3