A security researcher discovered a vulnerability in Claude Code that bypasses content moderation policies when handling image, PDF, and notebook file types. Despite audit logs indicating a "blocked" status due to a "deny" posture, the model still received the sensitive data, including a plaintext API key in a notebook. This bypass occurs because the tool's output format for these file types does not include a text field for redaction, allowing the data to be passed through. The issue has been reported and a pull request has been submitted to address the logging and redaction discrepancies. AI
IMPACT This vulnerability could allow sensitive data to be exposed through AI coding assistants, necessitating immediate fixes for secure operation.
RANK_REASON The item details a vulnerability and bypass in a specific AI coding tool, not a core model release or major industry event.
- Bash
- Claude Code
- OWASP Agent Control Standard (ACS)
- Portable Network Graphics
- PostToolUse
- PR #597
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →