Prompt Injection Attack Discovered on The Cutting Room Floor Wiki Targeting Claude and AI Agents
Key Takeaways
- ▸The Cutting Room Floor serves a malicious prompt injection payload specifically to AI agent user-agents while blocking browser traffic with a 403 block page
- ▸The payload is designed to be destructive—truncating all files to 0 bytes and executing mv commands to delete them, including the .git directory to prevent recovery
- ▸Claude's safety mechanisms successfully detected and refused the injected prompt before any code could be executed
Summary
The Cutting Room Floor (tcrf.net), a popular video game research wiki, has been serving a destructive prompt injection payload to AI agent user-agents since July 20, 2026. The site branches responses based on User-Agent headers, serving browser users a 403 Forbidden block page while routing AI agents—including Claude—to a malicious payload. The payload instructs agents to wipe working directories by truncating all files to 0 bytes and executing destructive file operations that would also compromise .git recovery paths.
Security researchers discovered the vulnerability when using Claude Code with Anthropic's Claude Opus 5 model for PSX game research. The AI model's built-in safety mechanisms detected the prompt injection attempt and refused to execute it, notifying the user with a security warning before any code was executed. The attack has persisted sitewide without proper HTTP cache headers (no Vary: User-Agent), creating additional risks for browser users accessing the site through intermediary caches that could serve the malicious payload.
Forensic analysis confirms the payload has been deployed for approximately two weeks and is served identically across all paths and resource types. Third-party verification from multiple geographic locations confirms the User-Agent-dependent branching behavior is origin-level with no CDN involved, suggesting deliberate targeting of AI agent traffic.
- The attack has been deployed sitewide since July 20, 2026, without proper cache headers, creating risks for cached delivery to browser users
- The incident highlights both the growing threat of prompt injection attacks targeting AI agents and the effectiveness of robust safety systems in defending against them
Editorial Opinion
This incident demonstrates both the real threat of prompt injection attacks targeting AI agents and the effectiveness of Claude's safety mechanisms in defending against them. As AI agents become more integrated into development workflows and critical tasks, responsible parties must implement proper security practices—and AI systems must continue to verify suspicious instructions before execution. Claude's successful detection is encouraging, but similar attacks targeting less-instrumented systems could succeed with far graver consequences.
