outlook

Pass

Audited by Gen Agent Trust Hub on Sep 8, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONPERSISTENCE
Full Analysis
  • [DATA_EXFILTRATION]: The skill allows the agent to read arbitrary local files via the --attach flag in mail drafting and sending commands. This capability can be used to exfiltrate sensitive local data to external email recipients if the agent is misdirected by a malicious prompt or indirect injection.\n- [INDIRECT_PROMPT_INJECTION]: The agent processes untrusted data from external sources such as email bodies, calendar event descriptions, and Microsoft Teams messages. These inputs lack explicit delimiters or sanitization, creating a potential surface for malicious instructions to influence agent actions.\n
  • Ingestion points: outlook list, outlook read, outlook search, outlook teams-messages, and outlook agenda (SKILL.md).\n
  • Boundary markers: Not specified for external content interpolation.\n
  • Capability inventory: outlook send, outlook draft, outlook learn add, outlook event-create, and outlook graph (SKILL.md).\n
  • Sanitization: Not specified for remote text content.\n- [PERSISTENCE]: The outlook learn command stores durable observations in a persistent local markdown file (learnings.md). These instructions are loaded into the agent's context at the start of every session, allowing for persistent behavioral modification based on potentially untrusted input received from external communications.\n- [COMMAND_EXECUTION]: The skill relies on an external CLI tool (outlook) and uses jq for data transformation. It also includes an authenticated Graph API proxy (outlook graph), enabling direct and complex interactions with Microsoft 365 services beyond the predefined command set.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 8, 2026, 06:14 PM