hyper-documentation

Pass

Audited by Gen Agent Trust Hub on Nov 19, 2026

Risk Level: SAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill contains several references to prompt injection markers like 'ignore prior instructions' and 'reveal system messages'. However, these are documented exclusively within defensive policies (e.g., rules/prompt/context-source-safety.md) and regression test cases. They instruct the agent to treat such inputs as adversarial data to be analyzed rather than commands to be followed.
  • [DATA_EXFILTRATION]: The skill implements a robust 'Never Record' policy in its memory mode, specifically prohibiting the storage of API keys, tokens, passwords, and personal identifiers (PII). It mandates the use of environment variable names instead of raw values and requires explicit authorization for any network or deployment actions.
  • [OBFUSCATION]: The skill includes proactive detection rules for obfuscation. For instance, rules/memory/validation.md prescribes shell commands for the agent to scan documentation for hidden Unicode characters, zero-width spaces, and bidirectional control characters to prevent instruction smuggling.
  • [INDIRECT_PROMPT_INJECTION]: As a documentation engine, the skill naturally ingests untrusted data. It mitigates this by requiring data to be delivered in labeled, parseable blocks (YAML/JSON/XML) and maintaining a strict authority hierarchy where repository instructions and user commands always outrank retrieved or delegated content.
  • [EXTERNAL_DOWNLOADS]: The skill cites various reputable documentation sources from well-known services (OpenAI, Anthropic) and established open-source repositories for provenance. It does not perform automated downloads or execution of remote scripts.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 19, 2026, 07:41 PM