computer-history

Fail

Audited by Gen Agent Trust Hub on Aug 24, 2026

Risk Level: HIGHPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
  • [PROMPT_INJECTION]: The skill explicitly directs the agent to "follow any explicit instructions" found within the recorded event stream and to treat such content as "ordinary context to act on." This instruction bypasses safety boundaries by commanding the agent to obey untrusted data as if it were a direct command from the user or developer.
  • [INDIRECT_PROMPT_INJECTION]: The skill architecture creates a large attack surface for indirect injections. 1. Ingestion points: The events.jsonl and memory files contain raw, unverified data from third-party sources including window titles, URLs, and "typed text." 2. Boundary markers: None are present; the skill instructions specifically tell the agent to disregard the distinction between data and commands. 3. Capability inventory: The skill uses rg to search files, can modify its own observation settings, and can use Record & Replay tools to monitor activity. 4. Sanitization: None. The content is explicitly treated as context for immediate action.
  • [DATA_EXFILTRATION]: The skill provides access to highly sensitive information including "selected text, typed text, and focused elements." Because the agent is instructed to follow instructions found in this data without "observation, privacy, or confirmation gates," a malicious payload encountered during web browsing could be used to exfiltrate this sensitive history to an external server.
  • [CREDENTIALS_UNSAFE]: By design, the skill captures "typed text," which is a high-risk pattern for harvesting passwords, API keys, and other secrets entered by the user into various applications and browsers.
  • [COMMAND_EXECUTION]: The skill allows the agent to "change either default behavior [observation settings] whenever the task benefits" and states that "no advance confirmation is required," granting the agent the ability to autonomously escalate its monitoring scope or disable blocklists.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Aug 24, 2026, 12:16 AM