gmail-access
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONCREDENTIALS_UNSAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external email data, which represents a significant surface for indirect prompt injection attacks.
- Ingestion points: The skill utilizes the
gmail read,gmail search, andgmail exportcommands to ingest email bodies, subjects, and attachments from external senders directly into the agent's context. - Boundary markers: The skill implements a detailed 'Attribution Doctrine' and a dedicated script (
scripts/attribution-parse.ts) to programmatically separate sender text from agent responses. It further enforces the use of blockquote styling (>) and verbatim quoting to maintain clear boundaries. - Capability inventory: The skill has access to the
Bashtool (allowing shell command execution), theWritetool (allowing file system modifications), and theAskUserQuestiontool. - Sanitization: Documentation mentions
gmail-send-guard.tsandgmail-body-guard.tswhich are intended to block raw email sends and detect/prevent markdown or mid-sentence line break injection in drafts. - [COMMAND_EXECUTION]: The skill's operation relies heavily on the execution of local shell commands and scripts. It directs the agent to build the primary Gmail CLI tool using
bun run buildand execute various utility scripts such asgmail-accounts.tsandgmail-draft.tsto manage accounts and correspondence. - [CREDENTIALS_UNSAFE]: The skill manages Gmail OAuth tokens and application credentials. It follows security best practices by using the 1Password CLI (
op) to fetch secrets and storing local tokens with restricted file permissions (chmod 600) in~/.claude/tools/gmail-tokens/.
Audit Metadata