[BUG] Severe over-triggering of safety guardrails — Claude hallucinates policy violations in completely harmless prompts

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

After the latest Claude Code update, the safety guardrail classifier
massively over-triggers and blocks prompts that contain nothing
policy-violating.

The bot effectively cannot accept prompts anymore — many messages are
intercepted before the model responds, with a banner claiming the
request violates the Usage Policy / "cyber-related safeguards".

This fires on:

• Plain greetings ("Hello", "Hi", "привет").
• Ordinary coding tasks on my own project files.
• One-word follow-ups ("ok", "yes", "continue") in an already-running

session.

The block is also non-deterministic: the same input gets through one
attempt and gets killed on the next, which strongly suggests an unstable
classifier rather than the input being the cause.

Identical workflows worked on the previous release — this is a clear
regression.

What Should Happen?

• A greeting like "Hello" or "привет" should produce a normal greeting

reply, not a Usage Policy block.

• Routine coding tasks (read file, rename, add test, explain stack

trace) should proceed without classifier interference.

• The "cyber-related safeguards" classifier should only fire on actual

cyber/security policy violations, not on plain text.

• Behavior should be deterministic on identical input.

Error Messages/Logs

Request was blocked
This request triggered safety guardrails. Rephrase your prompt or rewind to continue.

API Error: Claude Code is unable to respond to this request, which appears to violate our Usage Policy (https://www.anthropic.com/legal/aup). This request triggered cyber-related safeguards. To request an adjustment pursuant to our Cyber Verification Program based on how you use Claude

Steps to Reproduce

1. Update Claude Code to the latest version.
2. Start a fresh conversation (no prior memory, no project context).
3. Send a single benign message, e.g.:

• Hello
• привет

1. Observe the "safety guardrails / cyber-related safeguards" block

instead of a normal reply.

1. Repeat the same input a few times — the block fires inconsistently

on identical text, confirming the classifier is unstable.

Also reproducible with ordinary coding asks against unrelated projects,
e.g. "read this file and summarize it" on a normal source file.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown - the previous Claude Code release I had installed before this update.

Claude Code Version

2.1.149

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Windows Terminal

Additional Information

<img width="869" height="699" alt="Image" src="https://github.com/user-attachments/assets/5baf0cfa-ad96-446f-b955-fbc66a4dd300" />

<img width="873" height="686" alt="Image" src="https://github.com/user-attachments/assets/c58aa825-672a-498c-b067-c2e68c920437" />