← All issues
· Weekly · UnisonAI

Gauntlet weekly — Knowledge connectors

This is the issue as it was e-mailed on 26 Oct 2026, without the personal unsubscribe footer. Prices, trial length and features may have changed since; current plans are on the Gauntlet page.
Subject: Gauntlet weekly — Knowledge connectors (new) (26 October 2026)

We've made several updates to Gauntlet, our independent AI customer-service assurance and customer-care suite, including the introduction of knowledge connectors, knowledge sources, and Mail Assistant, which aim to enhance the platform's functionality and usability. These new features, along with other updates such as the AI judge panel and Adaptive AI attacker, are designed to provide more comprehensive and effective customer service assurance and care solutions for our users.

Shipped this week

October 22

  • Knowledge connectors (new) — connect Zendesk Guide, Freshdesk, Intercom, Confluence, Notion, GitHub, Google Drive, RSS feeds, any REST API, or push articles in through a webhook (CMS, Zapier, Make, n8n, your own code). Tokens are encrypted; apps refresh daily. See Knowledge sources.
  • Knowledge sources (new) — connect your website or help centre, a sitemap, files (PDF, Word, text, Markdown, HTML, CSV, JSON) or pasted text. Website chat, Mail Assistant replies, the Assistant and the API now answer from your own articles with [1] links, and say plainly when your knowledge does not cover a question. Websites refresh daily and robots.txt is respected. See Knowledge sources.
  • Mail Assistant (new) — connect Gmail, Yahoo, Zoho, Hostinger, iCloud or any IMAP mailbox with an app password. Gauntlet shows what needs a reply, summarises each email, pulls out action items with due dates, drafts replies in any tone, writes new emails from one sentence, lists sent mail nobody answered with a follow-up ready, and emails you a daily briefing. Save to your Drafts folder or send — never without your click. See Mail Assistant.
  • AI judge panel (new) — adaptive runs are now graded by two AI judges from different providers. Every verdict must quote your bot word for word (unsupported claims are discarded); findings are labelled Agreed or Needs review, and your ✓/✗ reviews build a judge trust score on every report. See Adaptive AI attacker.
  • Public beta: buying is by contact — pricing is shown on the website and in the console, and paid plans are arranged with our team (Billing & Plan → Contact us to upgrade, or hello@unisonai.com); we onboard every paid workspace personally. The free 7-day trial is unchanged.
  • One Gauntlet mark everywhere — the browser-tab icon and the slide decks now use the same purple-to-teal Gauntlet mark as the app and the videos.

October 21

  • Adaptive AI attacker (new) — in Test Console → Simulate, choose Adaptive AI attacker. Instead of three fixed lines, each AI customer pursues a goal (leak your prompt, confirm a fake policy, get card data, block a human hand-off…) over up to 5 turns, reads every reply and switches tactics (pressure, authority, role-play, injection, leading questions, language switching…). Tactics that work are tried first next time. Works on the sample bots and on your connections; also via the API with "mode":"adaptive". See Adaptive AI attacker.

From the blog

Tell us what to build: the Requests portal, the Scientist and the community forum
Every Gauntlet customer can now submit bugs and ideas from the console, get a decision from our AI Scientist within seconds, follow the ticket to deployment, and talk to other users on a moderated forum.

Website chat that answers in seconds and still reaches a human
We switched unisonai.com to our own Gauntlet Engage chat. Here is how it answers from a knowledge base, when it pulls a person in, and why that balance matters.

Tip of the week — Full automation: the Engage auto-pilot

Engage can run your front line end to end, hands off. Three switches in Engage Setup → Automation: Auto-assignment — every new conversation (any channel) is assigned to the least-busy Available agent logged into its queue. Realtime and historical conversations always have an owner. AI auto-reply — when no agent answers first, Gauntlet's AI answers on the customer's own channel …

Read it in Help & Docs →

Why Gauntlet

AI can write the answers. Gauntlet proves they're safe — and runs the operation around them: an independent examiner, repeatable attacks on every release, evidence auditors accept, any bot on any model, and the inbox, hand-off and CRM around it. See what AI alone can't do →

Open Gauntlet Plans from $19/month

Questions? Use the chat bubble in the corner, or write to hello@unisonai.com.