TheMailer phishing kit lets AI write, test and rewrite every email until it reaches the inbox
Varonis details TheMailer, a kit that plugs Claude, ChatGPT and other models into nearly every phishing step. Typos are no longer a tell; DMARC and click-time checks are.
Varonis Threat Labs has documented TheMailer, a phishing kit advertised to criminals since March 2026 that plugs Claude, ChatGPT and other commercial AI models into almost every step of an email campaign, from writing the lure to rewriting it until it lands in the inbox. Any organization that receives email is a potential target. The useful response is not new: enforce DMARC, check links at click time, and update awareness training, because polished language is no longer a warning sign.
What happened, in plain English
A phishing kit is ready-made software that criminals buy so they don't have to build their own scam emails, fake pages and sending setup. On November 17, 2026, the security company Varonis published research on one called TheMailer. According to Varonis, it is advertised on a cybercrime network for $3,000 to $6,000 for a lifetime license, paid in cryptocurrency. The buyer brings their own AI account (an API key, the password a program uses to call an AI service), their own mail servers and their own list of targets.
What makes TheMailer different is how much of the work it hands to a large language model (LLM), the kind of AI behind chatbots. The operator types a short description, or imports a real email from a brand, and the AI writes the message. Each recipient then gets a slightly different email, subject line and attachment. Before the real send, the kit mails a test to a seed mailbox it controls, checks whether the message landed in spam, and asks the AI to rewrite it and try again until it gets through.
Think of a scammer who used to photocopy one letter and mail it to thousands of people. Now he has an assistant who writes every letter by hand, tests each version against the post office's sorting machine, and changes the wording until it gets through. Spam filters that learn to spot "the letter" have nothing fixed to recognize, and the old advice to look for spelling mistakes stops working. As Varonis puts it: "Bad grammar isn't a reliable sign anymore."
Are you affected? What to do now
Everyone who runs or uses business email is in scope. Varonis does not name confirmed victims, and its figures come from the seller's own screenshots: one shows a campaign of 24,829 emails, another a list of 145,133 targets, almost all of them Japanese company and university domains (.co.jp and .ac.jp). Varonis also saw signs of lures themed on financial services, account security and hotel bookings, and targeting of German, French and Russian-language mailbox providers. One sample subject line impersonated Robinhood.
We found no list of indicators of compromise (IOCs, the addresses or file fingerprints defenders can block) in the Varonis write-up. Because every email and attachment is unique by design, fixed blocklists would help little anyway. The defenses that work are the ones that check where a message came from and where a link really goes.
Checklist for IT and security teams
- Lock down your own domain. Make sure SPF, DKIM and DMARC are set up, and move DMARC to
p=rejectonce reports show legitimate mail passes. (SPF and DKIM let receiving servers check that mail claiming to be from you really is; DMARC tells them to reject mail that fails.) Varonis recommends exactly this. - Check how your mail gateway treats inbound failures. Confirm it actually honours senders' DMARC policies rather than only tagging failures.
- Watch for lookalike domains. Monitor newly registered domains that resemble yours, your suppliers' and your banks'.
- Use click-time link checking. Varonis advises following the full redirect chain when a user clicks, including links hidden behind Google and other trusted domains. TheMailer uses such "trust redirects".
- Treat attachments as a delivery route. The kit builds a unique PDF per recipient (for example "Invoice_John.pdf"), SVG image files that redirect the browser, and calendar invites. Review whether your gateway inspects SVG files and calendar invitations from outside senders, and block SVG attachments if your business doesn't need them.
- Update awareness training this quarter. Drop "look for typos" as a key sign. Teach staff to pause on urgency, unexpected requests, "unusual activity" alerts and any login prompt reached from an email, and to report rather than reply. The kit has a module that collects replies and can answer them with AI, so a reply may start a conversation with the attacker.
- Limit the blast radius. Phishing-resistant multi-factor authentication (MFA) such as passkeys or security keys, plus least-privilege access, limits what one stolen password can unlock. Varonis also notes that "least-privilege access and data monitoring limit the damage if someone clicks anyway."
If your domain already enforces DMARC at reject, your gateway does click-time scanning and your training was refreshed recently, there is nothing urgent to change. This is a reason to check those controls, not a new emergency.
The expert view
Using LLMs to write phishing text is old news; it has been the most-cited criminal use of generative AI since 2023. What Varonis describes is a step further: the model sits inside a feedback loop. The "Inbox Loop" sends a test message, checks placement over IMAP (the protocol for reading a mailbox), and asks the model to revise content when the message is filtered. An "Inbox Finder" tries combinations of sending accounts, headers, encoding and DKIM settings against a seed mailbox. "AI Autopilot" groups recipients by mailbox provider and feeds the model provider-specific notes on spam filtering. An "AI WarmUP" feature reportedly runs send-and-reply exchanges so fresh sending accounts look established. In effect, the attacker gets automated A/B testing against the same filters defenders rely on.
The defensive consequence is the one Varonis highlights: with per-recipient variation, "there's no single template to block." Content signatures and hash-based attachment blocking lose value, since the kit also randomizes PDF metadata so no two files share a hash. The controls that keep their value are the ones attackers cannot reword: sender authentication, domain reputation and age, redirect-chain analysis, and identity controls after the click. Dark Reading quotes Varonis researcher Daniel Kelley calling the result "opportunistic phishing with the polish of something targeted". The kit personalizes at scale, but it doesn't research individual victims the way a spear-phishing operator would.
Two details are worth noting for the AI providers' side of the story. First, the kit is model-agnostic: Varonis lists Claude, ChatGPT, Grok, Gemini, Mistral, DeepSeek and the OpenRouter gateway. Banning one provider's accounts doesn't stop it. Second, Varonis says the tool's configuration page notes that "Anthropic and OpenAI block requests from cloud IP ranges" and advises routing traffic through SOCKS5 or residential proxies. That suggests provider-side abuse controls are adding friction, and that criminals are working around them. We found no public comment on TheMailer from Anthropic or OpenAI.
This fits a pattern seen across recent threat-intelligence reporting. Earlier this month, SOCRadar described a Russian-speaking actor whose XaocMail ClickFix campaign also generated close to a thousand personalized lure pages, with tooling that SOCRadar says carries artifacts of development with Claude Code. AI is lowering the cost of polish, variation and iteration, not inventing new attack classes.
What is still unknown: Varonis does not say whether it ran the software or observed live campaigns. The campaign sizes come from seller screenshots and are unverified. The number of buyers and the success rate are not known, although Varonis assesses that the customer base is large. No CVEs are involved, and no attribution to a specific group or country has been made.
Official sources
- Varonis Threat Labs: "Meet TheMailer: The AI Phishing Kit that Rewrites Emails and Taps Top Models" (November 17, 2026), the primary research
- SOCRadar: XaocMail ClickFix campaign (November 13, 2026), related AI-assisted phishing operation
- Coverage: Dark Reading: "'TheMailer' Phishing Kit Puts AI in the Driver's Seat"
- Coverage: GBHackers: "New TheMailer AI Phishing Kit Uses ChatGPT and Claude to Bypass Email Spam Filters"