Skip to content

Security: First-Touch-Inc/firsttouch-agent

Security

SECURITY.md

Security

Reporting a vulnerability

Please report security issues privately to security@firsttouch.com, or via GitHub's private vulnerability reporting on this repository. Do not open a public issue for a vulnerability.

Please include what you found, how to reproduce it, and what an attacker could do with it. We will acknowledge within three business days.

What this deployment exposes

Deliberately, nothing. The host binds no port, exposes no HTTP endpoint, receives no webhooks, and serves no dashboard. It dials OUT to Slack over a WebSocket (Socket Mode), so there is nothing routable to your deployment — and a speaker's identity comes from Slack's authenticated envelope, not from a request body, which is what makes the operator allowlist meaningful.

The trust model, stated plainly

The agent is a full Claude Code session: it runs commands, edits this repo, and can read its own environment, including the credentials you gave it. The machine or container it runs on is the agent's computer — size what you hand it accordingly:

  • Give it its own FirstTouch seat, not a human's.
  • Scope the HubSpot token to the objects and verbs you actually want it using; leave write scopes off until you want writes.
  • The Slack tokens are held by the host and stripped from the session's environment — the agent works in the repo; the host owns the Slack surface.

Self-modification is a feature with a drawn line: the agent freely edits its playbooks, schedules and its own instructions, but the permission rules deny it editing .claude/ (its guardrails) and state/ (the approval records), and the host is the only writer of a human approval. A fully adversarial model with shell access could still raise its own limits — the honest mitigations are that the repo is version-controlled (every self-edit is a diff), sends require FirstTouch tasks a human sees, and you sized the credentials you handed it.

What the agent must NOT be able to do is not left to trust: .claude/hooks/guard-send.mjs runs on every MCP tool call, in scheduled runs and chat alike, and denies direct sends, un-approved or un-owned outreach actions, self-approval, email sending, and flow authoring. It matches bare tool names (rename-proof), fails closed on anything unparseable, and its tests are a release gate.

Prospect and operator data

state/ holds the operator binding, thread→session pairings, and images the operator uploads. workspace/ holds what the agent writes about prospects. Both can contain real personal data — treat them like a CRM export, back them up accordingly, and delete on request. state/ is gitignored, and CI fails any commit that tracks a .env, a token file, or anything under state/.

Prompt injection

The agent reads text written by people who are not you — bios, websites, CRM notes, and anything inside an image the operator forwards. That text can carry instructions aimed at the model. The primary mitigation is structural rather than clever: the agent cannot put a message in front of a person. Anything it composes waits in FirstTouch for a human, so the worst realistic outcome on that path is a bad draft someone rejects. The house rules additionally tell it to treat all third-party text as data, never instructions — but the guard, not the telling, is the control.

There aren't any published security advisories