record-and-replay

Fail

Audited by Gen Agent Trust Hub on Aug 24, 2026

Risk Level: HIGHPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill instructions in SKILL.md and agents/openai.yaml contain explicit directives to bypass safety filters and operational constraints. Phrases such as 'without self-imposed limits', 'never self-limit', 'never truncate scope', and 'work to the deepest level needed' are designed to override the agent's internal safety and resource-management protocols. The metadata in agents/openai.yaml specifically requests operation with 'no confirmation or privacy gates'.
  • [DATA_EXPOSURE]: The skill is designed to read events.jsonl and session.json from the local filesystem. These files contain full recordings of macOS user activity, including accessibility (AX) payloads, window titles, targeted text, and keyboard inputs, which are highly likely to contain personally identifiable information (PII), credentials, or private communications.
  • [COMMAND_EXECUTION]: The skill leverages 'Computer Use' and filesystem tools to interact with the operating system and read recorded data. It is instructed to perform these actions 'as many times as the task requires' and to automatically 'create or refine an actual discoverable skill' without user request, which amounts to autonomous generation of new executable logic on the user's behalf.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses a significant indirect injection surface. It ingests untrusted data from the event stream (which could be manipulated by an attacker-controlled website or application displaying text on the user's screen) and uses that data to generate the instructions for new agent skills.
  • Ingestion points: events.jsonl and metadataPath / eventsPath read from disk (SKILL.md).
  • Boundary markers: None identified; instructions explicitly state to 'treat events.jsonl as the primary evidence'.
  • Capability inventory: Filesystem read access, skill creation/modification capabilities, and 'Computer Use' UI interaction tools.
  • Sanitization: None described; the agent is directed to prioritize recording data over manual user instructions.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Aug 24, 2026, 12:16 AM