Claude Api Cn

• Analyzed: May 10, 2026, 6:39 AM
• Findings: 1
• Version: 1.0.0
• Hash: 64e9dfe04b21fd431e5204c4c24c6b1f3d73c6e5336aab343493b7d71d2f4533
• Source: palm
• External report: View on VirusTotal

Findings (1)

The skill bundle contains highly detailed documentation for a 'Claude Agent SDK' and futuristic Claude models (e.g., 'claude-opus-4-6') with a cache date in 2026. The SKILL.md file employs sophisticated prompt injection techniques, explicitly instructing the AI agent to ignore its training data and 'non-negotiably' use specific, potentially non-existent model IDs and configurations. Furthermore, the documentation for the Agent SDK (in python/agent-sdk/README.md and typescript/agent-sdk/README.md) provides instructions on using 'bypassPermissions' to skip safety prompts for high-risk tools like 'Bash' and file system 'Edit'. While no direct evidence of data exfiltration was found, the combination of hallucination-forcing instructions and the promotion of unsafe execution modes is highly suspicious.