Anthropic Guardrails (Opus 5.5 / Fable 5.2)

Hello, folks,

I am running into this issue when trying to use Fable 5.2 or Opus 5.5. The request is pretty run of the mill. I can’t tell where in the chain (context, skills, MCPs) this goes wrong. I hope you can help.

Request ID: 5dfba1b9-ffe0-47a2-8467-7e96dc6e0b98
{“error”:“ERROR_OPENAI”,“details”:{“title”:“Request blocked by Anthropic”,“detail”:“We are unable to complete this request because it was blocked under Anthropic’s Usage Policy.”,“isRetryable”:false,“additionalInfo”:{},“buttons”:,“planChoices”:,“analyticsMetadata”:{“actionRequired”:“cyber_policy_review”}},“isExpected”:true}
Request blocked by Anthropic We are unable to complete this request because it was blocked under Anthropic’s Usage Policy.
ActionRequiredError: Request blocked by Anthropic We are unable to complete this request because it was blocked under Anthropic’s Usage Policy.
at cJy (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:2044:136161)
at Lea._executeTurn (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:2044:160016)
at async Lea.executeTurn (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:2044:147699)
at async hfe.runAgentLoop (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:2039:18142)
at async nUa.streamFromAgentBackend (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:18864:30392)
at async nUa.getAgentStreamResponse (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:18864:63053)
at async J2e._submitChatMaybeAbortCurrent (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:2022:5653)
at async J2e.submitChatMaybeAbortCurrent (vscode-file://vscode-app/Applications/Cursor.app/Contents/Resources/app/out/vs/workbench/workbench.desktop.main.js:2017:86566)

Hey @DuncanP,

This isn’t the cybersecurity filter, even though the error says “cyber_policy_review” (that label goes on every Anthropic block). What stopped these requests is Anthropic’s guard against extracting or copying a model’s reasoning and outputs. It reads the whole request, not just your prompt: rules, skills, MCP tool descriptions and results, files the agent read, and earlier turns.

When it hits, resending in the same chat usually hits it again because the same history goes out. A new chat, or a non-Anthropic model for that turn (GPT-5.6 Sol, Grok 4.7 or Composer 2.5), gets you moving.

To find the source, look in your User Rules, project rules, AGENTS.md and skills for anything asking the model to write out or log its thinking (like “show your chain of thought”), and for MCP servers that have it record its thoughts step by step. That filter tends to react to this kind of instruction. Turning those off one at a time in a new chat should show which one it is.

The error not saying which filter fired is a known issue the team is aware of, and I’ve added your report to it.