AI is Making Spam Cheaper, Can We Replace Open Discord Invites with a Canister Gate?

Hi community,
here’s an idea.With AI lowering attack costs, cyber warfare is accelerating.

For example, spam and scam accounts enter public communities via open invite links, post freely, and flood members and admins with DMs.What if we block public invite links entirely and route all entry through a single canister-based approval gate?

Proposals, consultations, and PR submissions could also go through the same system, using agent micropayments.At the moment,
the main risk I see is an attack that floods the canister with a large number of update calls, draining cycles.

I’ll try building a prototype over the weekend and will post progress here.

Progress update: it exists, and the interesting part isn’t the security

I said I’d build a prototype over the weekend. Here’s what exists — and while it started as a spam defense, the thing I most want to compare notes on now is using Internet Identity as ordinary community-operations infrastructure.

What it does

One page, one form. You pick why you’re here — join, send a proposal, book a paid consultation, or submit a PR listing — write a few lines, and sign in with Internet Identity at the moment you submit, not before. The application lands in a canister. An admin reviews it from a browser panel. If you’re approved, a one-time Discord invite appears on the page you applied from: nothing is emailed or DMed, so there’s no link in transit to phish or forward. (It’s fetched with an update call, not a query — query responses aren’t certified, and this is the one link a user will actually click.)

For the paid lanes the canister holds an ICRC-2 deposit and returns it on approval, on rejection, or automatically via timer if nobody reviews it in time. The only path where you lose it is an explicit spam judgement, and the count and total are published.

Why Internet Identity and not a normal form. A Google Form would have taken an afternoon. But then I’m collecting email addresses — which are free to generate, so the spam problem just moves one step back — and I’m holding personal data I don’t want to hold. II gives me the inverse: nothing stored, nothing to leak, no address to burn through, and the same identifier doubles as the refund destination once money is involved. The applicant proves they are a persistent identity without telling me who they are. For a community intake desk that’s a better trade than any form service I could buy, and I don’t think it’s a use of II I’ve seen written up anywhere.

One decision I’d repeat: the admin panel can’t move money. Review, invites, and reversals are gated on an Internet Identity admin roster. Withdrawals, prices, and adding admins need the controller identity, which only lives in dfx. A compromised browser session can admit the wrong person; it can’t take the treasury.

Not on mainnet yet — 59/59 locally, including an upgrade test. Free lanes go first; paid lanes stay switched off until someone reviews the money paths.

What I got wrong

I said the main risk was update calls draining cycles. That was the easy part (inspect + an O(1) counter). Three things turned out to be worse: an atomicity bug in the refund path that let confiscate and refund both succeed on the same deposit, a sybil attack that could lock everyone out of the free lane, and the fact that a Discord bot token can’t safely live in a canister at all. All three are dealt with — happy to expand on any of them in replies.

My question

Has anyone built a plain form around Internet Identity — an application form, an intake desk, anything like that? Not a wallet, not a dapp login. Every II example I can find is one of those two, and I’d like to compare notes with someone who used it for the boring operational side of running a community: did people bounce at the II screen, and did you keep it?

For context, since it’ll be obvious anyway: I run the community, I’m not a systems programmer, and this was built with heavy AI assistance. A year ago it would have been a spec handed to someone else, or more likely a Google Form.

Contact forms have become an entry point. Status update, and the first real application

Bottom line

A door that is free, open to anyone, unlimited, and always read by a human can no longer be defended. You can’t put a price on a company’s contact form. You can put one on a community’s entrance, so that’s where I’m testing it: the Discord of a community I run.

The problem

Every company website has a contact form, and reading every message that comes through it is someone’s job. That is exactly what the attacker counts on. Sending was already nearly free in 2021. What AI made cheap is the part that used to need a person: a message written for you, and a two-week conversation.

Real cases

Attacks that start at the contact form have been running for at least five years. The lure has barely changed; what was new in 2025 was a crew willing to talk for two weeks first.

  • April 2021, Microsoft. Automated submissions threatening legal action over “stolen images”, apparently bypassing CAPTCHA. Payload: IcedID, which Microsoft describes as an entry point for human-operated ransomware.
  • February 2024, Proofpoint. The same copyright threat, sent through contact forms. Payload: Latrodectus, likely written by the IcedID developers.
  • August 2025, Check Point Research. Contact opened through the form, a businesslike exchange kept up for as long as two weeks, then a ZIP “for NDA review” with a backdoor inside. Dozens of manufacturers; over 80% in the US, Japan included.

On damage: loaders like these feed human-operated ransomware crews, and the largest of them in 2021, Conti, extorted at least $180 million (Chainalysis). How much of that started at a contact form, no one has measured.

Then in October 2026 the inquiry records themselves became the target. An unauthorized party logged into the admin console of i-ask, a Japanese inquiry-management SaaS; up to 713,126 inquiry records from up to five client companies may have leaked, including the text of the inquiries. A price per attempt wouldn’t have stopped that one. I include it because the form’s archive is now a target in its own right.

The numbers

  • 53% of web traffic in 2025 was automated: 40% bad bots, 13% benign bots, 47% human (Imperva, 2026).
  • 12.5x. AI-driven bot requests blocked per day on Imperva’s network went from 2 million to 25 million (Thales/Imperva, 2026).
  • $20.9 billion in US internet-crime losses reported to the FBI in 2025, with an “AI-related” descriptor tracked for the first time: 22,364 complaints, $893 million.

The experiment

The gate is a Motoko canister with Internet Identity login.

  • The public invite link is gone from our site; joining goes through an application page. Discord’s own screening limits what members can do once inside; the invite link itself stays public and forwardable, and that’s what I wanted closed.
  • Applicants sign in with Internet Identity (ICP’s free, passkey-based login) and a human reviews each application.
  • Live on mainnet since late August 2026. The free application types (join, propose) are on; the hold-and-refund paid type from the prototype is deployed but disabled.
  • In seven weeks, one application has arrived. It was LLM-style boilerplate. I read it and rejected it.

That’s as specific as I’ll get. The rest would disappear from the next application.

What I learned

  • Internet Identity did its job: a stable pseudonymous principal I can attach a verdict, and later collateral, to. It doesn’t say a person is behind it, and it isn’t meant to; a new anchor is free.
  • While applying is free, a rejection costs the attacker nothing. It costs me cycles.
  • As models improve, applications will come with plausible histories and portfolios, not just plausible text. Judging by the writing won’t last. What remains is the cost per attempt.

Next steps

  1. Put a price on applying. The applicant approves the gate to take a set amount of ICP, and nothing moves unless the verdict is spam (ICRC-2 approve). The allowance can be revoked before the verdict, so this is a deterrent, not an escrow; I’m taking that trade because holding funds would walk into Japan’s payment-services regulation. This is the next build.
  2. Ask for a public footprint. GitHub, shipped work, a public track record. These can be faked, but a public account with history takes time to build, and time is cost.

Token gating (Collab.Land, Guild) is the obvious prior art. The difference here: nothing is held, collateral is taken only on a spam verdict, and for now that verdict is mine.

Both look like a bad fit for an open door: ask someone who found ICP last week for public credentials, and you may turn away exactly the people you want most. I think that gets it backwards. If applying stays free, solicitations and bots fill the queue, the human review breaks first, and the door is closed to everyone. A mechanism where a genuine applicant pays nothing and only the attacker pays is what keeps it open.

I’ll report back once the paid type is live.

What I’d like to hear

Has anyone put token gating, a cost, or a public-identity requirement in front of a community entrance? What happened to genuine sign-ups afterwards? One line of experience is enough. If you moderate the DFINITY Discord or this forum: what does your queue look like?

I run the community, and I built the gate myself with AI assistance. The code isn’t public yet, and I’m not naming the community here. If anyone here reviews Motoko, I’ll share both privately.

Sources: Microsoft, Apr 2021 · Proofpoint, Apr 2024 · Check Point Research, Aug 2025 · Chainalysis via The Record · i-ask incident, INTERNET Watch · Imperva Bad Bot Report 2026 · Thales/Imperva via Help Net Security · FBI IC3 2025 report (PDF)