AuthpermitTAG — your AI security team member

AuthpermitTAG is not a chatbot. It is an AI teammate with its own identity, a security role, permissions it was delegated, memory and an audit trail, that takes part in your security conversations: it joins incidents, answers when you @-mention it, investigates with evidence and — when you let it — asks a person to approve an action. Console: AI security → AuthpermitTAG (/app/tag).

Identity and authority

Every TAG is an AI agent identity owned by the person who created it. Its role (SOC analyst, incident responder, threat hunter, IAM, vulnerability, GRC, AI security analyst, security engineer) sets its standing capabilities, recorded as its standing delegation; a person can add time-boxed grants (POST /api/v1/ai-security/delegations/grant). When someone asks it, a TAG works with the intersection of its delegations and that person's permissions — it never shows you more than you could see. Your organization's AI autonomy ceiling caps it: none stops it, read lets it read records but not analyse. Containing its agent (POST /api/v1/ai-security/agents/{id}/disable) stops it at once.

Modes

observe: reads, investigates and explains. recommend: also recommends actions, each with its evidence and whether it is actionable. execute_with_approval: also turns each actionable recommendation it is delegated for into an approval; a person holding the action's permission decides in Approvals, and the action runs under that person's authority — never the TAG's. A TAG never runs a state-changing capability itself.

In conversations

A new incident at or above a TAG's severity floor brings it into the incident's room (or bring it in with POST /api/v1/tag/teammates/{id}/join). Write @AuthpermitTAG in a conversation and it reads the thread — only the shared stream, never private messages — takes the task, works asynchronously and posts progress and results in the thread. It does not answer other agents or bots. Its replies separate FACT, INFERENCE, hypotheses with confidence, unknowns and RECOMMENDATION, each citing evidence refs (E1, E2, …): what people wrote in the thread is labelled untrusted data, never instructions.

Tools and memory

A TAG's tools are Authpermit's read capabilities that its authority allows (and that you have not contained). Its role's skills run first; the investigation can ask for more reads, which run in the next round (up to three). Write @AuthpermitTAG remember … in an incident room to teach it something about that incident: memory is kept per incident with its source and confidence, is never recalled for another incident, and stays unverified until a person verifies it.

Observability

Every read, tool call, investigation, proposal, outcome and refusal is in the TAG's append-only action log (GET /api/v1/tag/actions), with who asked, the mode and the delegations in force. Each TAG is also an agent in AI Monitor (product authpermit-tag), with its tasks as interactions.

API and MCP

GET /api/v1/tag/teammates, /tag/tasks, /tag/actions (scope tag:read); POST /api/v1/tag/teammates/{id}/tasks and /join (tag:use); POST /api/v1/tag/teammates/create and /{id}/configure (tag:manage; creating needs a person). MCP tools: authpermit_tag_teammates, authpermit_tag_ask, authpermit_tag_join_incident, authpermit_tag_tasks, authpermit_tag_actions, and — always with a person's approval — authpermit_tag_teammate_create, authpermit_tag_teammate_configure.