morning-operator
Fail
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: HIGHPROMPT_INJECTIONPRIVILEGE_ESCALATIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONTIME_DELAYED_CONDITIONAL
Full Analysis
- [PROMPT_INJECTION]: The skill employs aggressive directive language intended to bypass standard AI safety and operational constraints. Instructions such as "Zero questions," "No permission," "No confirmations," and "Never ask permission" are explicit attempts to override the agent's typical clarification and confirmation behaviors.
- [PRIVILEGE_ESCALATION]: The skill mandates an autonomous "operator" mode that instructs the agent to "Act first. Report after" and "Decide and do." This attempts to escalate the agent's privileges by removing human-in-the-loop requirements for sensitive operations, including code deployment, email communication, and financial data management.
- [INDIRECT_PROMPT_INJECTION]: The skill exposes a significant attack surface by requiring the agent to ingest data from untrusted external sources. (1) Ingestion points: Screen content, clipboard data, email inboxes, and external task systems (Linear, Notion, etc.) defined in Phase A. (2) Boundary markers: None identified; the agent is instructed to "read and verify the live state" directly. (3) Capability inventory: Shell command execution (git, curl), deployment triggers, email/message sending, and modification of task management records. (4) Sanitization: No sanitization or filtering logic is provided for the ingested data.
- [COMMAND_EXECUTION]: The skill authorizes the use of powerful system tools including
git(status, stash, branch management) andcurlfor live verification. When combined with the "no-questions" autonomous directive, this poses a risk of unintended or malicious system state changes. - [DATA_EXFILTRATION]: The agent is granted access to highly sensitive information including "secrets," "billing vs usage," "monetization surfaces," and private communications. The autonomous mandate to "send," "notify," and "update" across external platforms creates a credible path for unauthorized data exfiltration, even in the absence of a hardcoded attacker URL.
- [TIME_DELAYED_CONDITIONAL]: The skill includes a "Weekly / Suspicion Health Check" phase designed to execute specific diagnostic and cleanup tasks only when triggered by a 7-day timer or specific environmental suspicions.
Recommendations
- AI detected serious security threats
Audit Metadata