Phishing-as-a-service just got a serious upgrade: a new kit called TheMailer uses large language models to run nearly the entire attack, from writing the lure to rewriting it until spam filters let it through, according to new research published Monday by Varonis.
The kit stands apart from the AI-assisted phishing tools security teams have grown used to. Where most kits lean on AI for one task — generating lure text, coding a fake login page, maybe automating repetitive steps — TheMailer automates much of the phishing process end to end, using whichever AI model the buyer chooses.
According to Varonis senior security researcher Daniel Kelley, the kit can tailor lures to each mail provider and “rewrite them until they bypass spam filters.” That second capability is the part defenders should sit up for: content filters, long the first line of email defense, are now the explicit target of automated iteration.
Phishing kits and phishing-as-a-service platforms have grown steadily more popular in recent years, and AI has made them more capable. Until now, that capability mostly showed up as lure generation, realistic phishing sites, reconnaissance, and simple task automation. TheMailer pushes further on both breadth — full-process automation — and depth — per-provider, per-recipient lure tuning.
A lifetime license, an API key, and go
The buyer’s journey starts with a lifetime license costing between $3,000 and $6,000 in cryptocurrency, with the price depending on the deal and which AI functions are included. The operator then plugs in their own API key and picks a model. Claude, ChatGPT, DeepSeek, Grok, Gemini, and Mistral are all supported.
From there, the buyer brings little more than infrastructure and a target list. Kelley said that customers supply their own SMTP servers, proxies, and target email lists; the kit handles the rest.
What the kit does with that list is where it diverges from bulk phishing of the past. TheMailer can generate lures from a short prompt or from a real brand email, personalize each message — email address, body, and subject line — strip out recognizable AI writing patterns, and adjust content to slip past spam filters.
Scale is built in. “The seller’s screenshots show a single campaign of nearly 25,000 emails and a list of more than 145,000 targets being sorted by mail provider. What’s different is that every recipient gets their own version of the email, subject line, and attachment, so a bulk campaign no longer looks like one,” Kelley said.
He drew a careful line, though, between personalization at scale and true targeting. “I wouldn’t call it a spear-phishing tool. The personalization we saw is at the level of a name on an invoice, and we didn’t see features for researching individual targets. It’s opportunistic phishing with the polish of something targeted,” Kelley said.
How effective is it? Nobody’s saying — yet
Varonis does not claim TheMailer has landed a confirmed success rate. Attacker-side LLM tools have historically been able to run tasks at scale, but their success rates vary, and no figure is attached to this kit. What the researchers do assess is that TheMailer has a large customer base.
Whether the kit works as advertised matters less than what it signals. Kelley expects copycats: “I’d expect more kits to follow, because once one seller offers this, buyers start asking for it elsewhere.”
That expectation reframes the threat for defenders. The old heuristic — bad grammar, awkward phrasing, generic greetings — loses its value when an LLM writes every message and scrubs its own fingerprints. A bulk campaign that reads like individual correspondence defeats both statistical detection and the human eye’s pattern-matching.
Varonis’ recommendations follow directly from how the kit works. Since TheMailer is built to defeat content filters, the firm points to email authentication — SPF, DKIM, and DMARC enforcement — plus monitoring for lookalike and newly registered domains, and checking links through full redirect chains at the moment of the click rather than at scan time.
On the human side, the firm says phishing training should reflect what emails look like in a post-AI world. “Bad grammar isn’t a reliable sign anymore, so teach users to be wary of urgency, unexpected requests, and login prompts, especially ‘unusual activity’ or account statement alerts,” the blog post read. Least-privilege access and data monitoring, the researchers add, limit the damage if someone clicks anyway.
The open question is how quickly this becomes the norm. TheMailer is one kit with one seller’s screenshots behind it; its real-world performance is unmeasured, and its customer base is an estimate, not a count. What is documented is a template — end-to-end AI automation, per-provider filter evasion, per-recipient personalization — now on sale, with buyers already asking for it elsewhere.