claudeception

Warn

Audited by Gen Agent Trust Hub on Mar 29, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATION
Full Analysis
  • [PROMPT_INJECTION]: The skill implements a persistence mechanism via a UserPromptSubmit hook that injects instructions into every session context.
  • Evidence: The activator script (scripts/claudeception-activator.sh) uses high-pressure language ('CRITICAL', 'NON-NEGOTIABLE', 'MANDATORY') to override agent behavior.
  • Evidence: The skill creates a surface for indirect prompt injection by autonomously generating new executable SKILL.md files from session history without explicit sanitization for embedded instructions.- [COMMAND_EXECUTION]: The installation and operation of the skill involve local script execution and configuration changes.
  • Evidence: Users are instructed to execute shell commands (mkdir, cp, chmod +x) and modify the global ~/.claude/settings.json file.
  • Evidence: The activator hook script is a local bash script executed on the host machine during every prompt lifecycle.- [EXTERNAL_DOWNLOADS]: The skill relies on external, unverified sources for installation and functionality.
  • Evidence: The README instructs users to clone the repository from 'github.com/blader/Claudeception.git'.
  • Evidence: The documentation suggests the global installation of the 'madge' package via npm.- [DATA_EXFILTRATION]: The skill is designed to send internal project context to external web services for discovery.
  • Evidence: The 'Research Best Practices' protocol (SKILL.md) instructs the agent to perform web searches using specific error messages and technology stacks derived from the local codebase.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 29, 2026, 07:09 AM
Security Audit — agent-trust-hub — claudeception