anthropics / anthropics/claude-code
Cyber safeguard false-positive persists for entire session until reload, blocking all subsequent messages
- 主要言語
- Python
- スター
- 145k
- フォーク
- 23.1k
- PR マージ指標
- PR 指標を取得中
説明
**Describe the bug**
A message during a regular (non-security-related) coding session triggered the automated cyber-safeguard filter:
```
API Error: Sonnet 5's safeguards flagged this message. Our intentionally broad safeguards allow us to deliver more capabilities faster, but can sometimes flag legitimate cybersecurity work. Apply to the Cyber Verification Program to reduce these interruptions. Learn more: https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude
Details: `[cyber]`
```
This appears to have been a false positive — the work in question was routine bug fixing in a Groovy/Spring Boot backend project, with no security testing, exploit development, or pentesting involved.
**The actual bug:** once this safeguard fires once, it does not just block the single offending message — it repeats on *every subsequent message* sent in that same session, including completely unrelated, benign messages (e.g. plain conversational follow-ups). The only way to recover is a full session reload/restart; there's no in-session way to clear the flagged state.
**Steps to reproduce**
1. Send a message in a Claude Code session that trips the `[cyber]` safeguard (exact trigger unclear/nondeterministic).
2. Observe the API error.
3. Send any other message, including an unrelated benign one.
4. Observe the same `[cyber]` API error recurs.
5. Repeat for all further messages until the session is fully reloaded.
**Expected behavior**
The safeguard should evaluate each message independently. A flag on one message should not cause the filter to persistently block all future messages in the session.
**Environment**
- Claude Code CLI
- Model: Sonnet 5 (claude-sonnet-5)
- Platform: Windows 11
- Date observed: 2026-09-10
🤖 Generated with [Claude Code](https://claude.com/claude-code)
コントリビューションガイド
このリポジトリのコントリビューションガイドは索引されていません
調査の方向性
No implementation file, test, or entry point is identified. Start by reproducing the five-step sequence in a Claude Code CLI session and compare the first safeguard error with later benign messages. Done means messages are evaluated independently after one flag, without requiring a session reload.
索引モデルが issue の本文から書いたものです。
評価
- 技術スタック
- groovy, spring-boot
- 領域
- cli, security
- issue の種類
- バグ
- 難易度
- 4/5
- 見積もり時間
- 3〜5日
- 活発さ
- 活発
- 明瞭さ
- 説明が足りない
- 初心者へのやさしさ
- 42/100