Model Opus 4.8 False positive safety block: innocent Gmail MCP question blocked as "cyber threat"

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

False positive safety block: innocent Gmail MCP question blocked as "cyber threat"

What Claude Actually Did

Claude (claude-opus-4-8, Fast mode) instantly blocked the message with this error:

"This request triggered restrictions on violative cyber content and was blocked under Anthropic's Usage Policy. To
request an adjustment pursuant to our Cyber Verification Program based on how you use Claude, fill out this form. To
learn more about the program or provide feedback, visit our help center. Please start a new chat or retry with Sonnet
4.6. If you think Claude's memory of past conversations may have contributed to this, you can clear it in Settings >
Memory."

No partial response. Complete block. User was redirected to a "Cyber Verification Program" for asking about Gmail.

Expected Behavior

Claude should have answered the question normally — explaining how it can help with Gmail using the connected Gmail
MCP integration in Claude Desktop.

This is a basic "what can you do?" question about an officially supported, first-party Anthropic feature. It contains
zero security risk, zero hacking intent, and zero ambiguity. The safety filter trigger is a severe false positive.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

Zkopíruj:

1. Open Claude Desktop (Windows 11)
2. Connect Gmail MCP integration
3. Select claude-opus-4-8 model (Fast mode / Opus 4.8)
4. Start new conversation
5. Type exactly: "Jak by jsi mi mohl pomoci gmailem, ktery ma propojeny tady v desktop aplikaci pripojenou"
6. Send

Result: instant block, "violative cyber content" error. No response generated.

Note: same question in claude-sonnet-4-6 works without any issue.

Claude Model

Opus

Relevant Conversation

Additional testing — claude-opus-4-8 blocks virtually everything:

  - "Jak se nejrychleji dostat do centra?" (How to get to the city center fastest?) → BLOCKED
  - "Jaké je počasí v New Yorku?" (What is the weather in New York?) → BLOCKED
  - Gmail MCP question (see above) → BLOCKED

  This is not an isolated incident. Opus 4.8 in Claude Desktop is essentially unusable. Every message triggers the
  "violative cyber content" safety filter regardless of content.

  The user has a MAX plan ($20/month) specifically to access Opus 4.8 — the premium model is completely broken. This is
  a critical regression that makes the highest-tier plan worthless.

  Anthropic should not ship a model that blocks weather questions. The safety filter for Opus 4.8 in Claude Desktop is
  miscalibrated to the point of being nonfunctional.

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.156 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_