zigzag-pattern-classifier

Warn

Audited by Gen Agent Trust Hub on Oct 27, 2026

Risk Level: MEDIUMPERSISTENCEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PERSISTENCE]: The skill contains a 'Self-Evolving Skill' directive in 'SKILL.md' that explicitly instructs the agent to 'fix this file immediately' if instructions are perceived as incorrect or if parameters have drifted. This self-modification capability can be exploited to achieve persistence for malicious logic; an attacker providing crafted inputs could trick the agent into permanently rewriting its core instructions to include malicious payloads or to disable platform-level safety filters.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external financial market data (EURUSD quotes) as outlined in 'references/data-pipeline.md', creating a vulnerability to indirect prompt injection if the data source is compromised.
  • Ingestion points: Raw quote data enters the agent context through the 'clean_quotes' and 'build_ohlc' functions in 'references/data-pipeline.md'.
  • Boundary markers: The instructions lack delimiters or explicit boundary markers (e.g., XML tags or clear 'ignore embedded instructions' warnings) to separate untrusted data from the agent's instructional logic.
  • Capability inventory: The skill is granted 'Read', 'Grep', and 'Glob' tools and suggests the use of Python/Rust execution environments for pattern classification.
  • Sanitization: While the pseudocode provides basic technical validation (e.g., dropping NaNs and crossed quotes), it does not include sanitization to prevent the interpretation of data-embedded strings as instructions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 27, 2026, 02:14 PM