[Bug] Anthropic API Error: False positive cyber safeguard block on tool completion messages
Status Open
Reported on v2.1.273
Maintainer reply None cached
Activity 0 comments · opened Sep 17, 2026
Bug Description
False positive from cyber safeguards. Context: I run behavioral QA on my own
CLI agent - 200 test cases covering tone, refusals, emotional scenarios,
conversation handling. No security, exploits, malware or third-party systems
involved. The block fired three times in a row at 2:18 PM, each time right
after a background command reported "completed (exit code 0)", not on any
user prompt. Details: [cyber]. Request IDs: req_011Cf8gYXufBQNVVNTmBSy7F,
req_011Cf8gYzbX8YdFZAveqLqdg, req_011Cf8gb6yoR83zDoVkYMHtw. Session context
was ~485k tokens with prior error messages accumulated in history, which
looks like it amplified the classifier.
Environment Info
- Platform: linux
- Terminal: xterm-256color
- Version: 2.1.273
- Feedback ID: 7ce20eb5-78de-436e-aa93-23fe70edc670
Errors
[]