SAFETY

Security model and limitations

Understand which controls TekMail enforces, where detection helps, and where your agent still needs a deliberate trust boundary.

The policy boundary lives outside the model

A TekMail API key selects one configured agent, its From address, inbox, permissions, usage limits, approval settings, and assigned policies. The caller cannot switch to another agent or choose a different sender in a tool argument.

Outbound enforcement

Before a message reaches the delivery queue, TekMail checks send permission, recipient-domain rules, plan and rate limits, approval rules, content policies, and recipient suppressions. A held message remains pending until an authorized human reviews it.

Inbound defense in depth

Inbound controls combine sender-authentication results, suspicious-instruction heuristics, business-email-compromise signals, quarantine thresholds, sender allowlists, and optional human approval. Held or quarantined messages are intentionally omitted from the agent inbox until released.

No detector is perfect. Treat every email body and attachment as untrusted input. Risk screening can prioritize or trigger review, but deterministic permissions and approvals—not a classifier verdict—should protect consequential actions.

Customer responsibilities

  • Keep agent keys on trusted servers and rotate or revoke exposed credentials.
  • Give each agent only the send, receive, domain, and attachment access it needs.
  • Require human review for financial, legal, credential, deletion, or other high-impact workflows.
  • Validate business facts in your own system before an agent acts on instructions received by email.
  • Use stable idempotency keys when retrying sends after timeouts or client reconnects.

Scope of the audit trail

TekMail records mail-side events and review outcomes. It cannot attest to actions a downstream agent takes after reading a released message, so retain the coding tool or automation platform's own execution logs as well.