caveman-cowork
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFENO_CODEINDIRECT_PROMPT_INJECTION
Full Analysis
- [NO_CODE]: The skill consists entirely of natural language instructions (Dutch and English) defining a specific communication style. It does not include scripts, binaries, or configuration files.
- [SAFE]: The instructions are focused on reducing token usage by removing filler words and politeness formulas. It lacks any capability to access the file system, network, or external credentials.
- [SAFE]: The skill includes an "Auto-Helderheid" (Auto-Clarity) safeguard that explicitly instructs the agent to fall back to full sentences for security warnings, irreversible actions, and complex instructions where brevity could cause ambiguity.
- [INDIRECT_PROMPT_INJECTION]: The skill evaluates user dialogue to determine when to activate its concise response style. While this represents a surface for indirect prompt injection, it is considered safe due to the lack of exploitable capabilities.
- Ingestion points: User chat inputs monitored for activation keywords such as "kort", "caveman", or "/caveman-cowork" (SKILL.md).
- Boundary markers: Not explicitly defined for the input data.
- Capability inventory: None detected. The skill uses no tools, scripts, or network calls.
- Sanitization: None present; the skill relies on the underlying model's safety guardrails.
Audit Metadata