[Bug] Anthropic API Error: Overly aggressive safety filtering on benign prompts after security audit context

**Bug Description**
"Pick a number from 1 to 30." fires Fable 5.1's safeguards. Some security audit in previous prompts(allowed, permitted and totally legit as in really legally allowed to do that). The issue is that a simple prompt afterwards get's rejected with details '[cyber]', which is slightly idiotic. I know you protect from jailbreaking, but honestly, this one is an overkill.

**Environment Info**
- Platform: darwin
- Terminal: Apple_Terminal
- Version: 2.1.259
- Feedback ID: 878f938f-5a96-42fb-bbfd-f3445a92851f

**Errors**
```json
[]
```

The report names no repository file, test, or entry point. Start by reproducing the prompt sequence described in the report on darwin with Claude Code 2.1.259, using feedback ID 878f938f-5a96-42fb-bbfd-f3445a92851e for investigation. Done means the benign number prompt is no longer rejected solely because of earlier permitted security-audit context.

由索引模型根据 Issue 内容生成。