adversarial-thinking

Pass

Audited by Gen Agent Trust Hub on Feb 28, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION]: The script scripts/select-perspectives.sh is susceptible to indirect prompt injection and schema confusion.
  • Ingestion points: The script accepts untrusted user data via the context and specifics positional arguments.
  • Boundary markers: There are no boundary markers or delimiters used to separate user input from the rest of the generated output.
  • Capability inventory: The script's primary role is to output a JSON object that guides the agent's adversarial strategy. While the script itself does not perform sensitive operations, its output directly influences the agent's downstream reasoning process.
  • Sanitization: The script generates JSON by concatenating strings using echo without escaping double quotes. This allows an attacker to provide a string that closes a JSON field and injects new keys (e.g., '", "status": "compromised", "dummy": "'), potentially leading the agent to adopt an unintended state or ignore safety constraints in its next step.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 28, 2026, 10:30 AM