# AI agents signing up for accounts

**URL:** <https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624>\
**Category:** Community Building\
**Created:** [13 באוקטובר,‏ 2026,‏ 10:44pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624 "2026-10-13T22:44:25Z")\
**Posts on this page:** 14\
**Page:** 2

<div class="post-metadata">

**Author:** ![Michelle\_L](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/michelle_l/32/484912_2.png) [@Michelle\_L](https://meta.discourse.org/u/Michelle_L)\
**Post date:** [13 בנובמבר,‏ 2026,‏ 10:22pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/22 "2026-11-13T22:22:09Z")

</div>

Are you using AI Spam detection or AI Triage to add additional guidelines around these types of posts to flag them for review by your team before posting to the community?

---

<div class="post-metadata">

**Author:** ![merefield](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/merefield/32/176214_2.png) [@merefield](https://meta.discourse.org/u/merefield)\
**Post date:** [13 בנובמבר,‏ 2026,‏ 10:25pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/23 "2026-11-13T22:25:03Z")

</div>

I have my own bespoke AI post analytics, but tbh, the posts are very human in length and structure. None of the usual “over formatting” or gratuitous length that AI posts sometimes have. There was only one with an em-dash.

None of the posts could be regarded as “spam” - they were on topic, thoughtful replies but somehow “out of the blue”.

It was the consistent patterns of their sign-up dates, usernames, geographical location and email addresses that bound them together and crystalised for me that they were suspect.

---

<div class="post-metadata">

**Author:** ![Michelle\_L](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/michelle_l/32/484912_2.png) [@Michelle\_L](https://meta.discourse.org/u/Michelle_L)\
**Post date:** [13 בנובמבר,‏ 2026,‏ 10:31pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/24 "2026-11-13T22:31:32Z")

</div>

We have seen some of that too, we do use AI Triage and it tends to flag the majority of the posts for us to review, even if they appear spot on. We also review New Registrations on a daily basis of accounts activated and not activated so that we can proactively catch some of those accounts that like to lie in wait as well.

---

<div class="post-metadata">

**Author:** ![Andrew\_Rowe](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/andrew_rowe/32/445877_2.png) [@Andrew\_Rowe](https://meta.discourse.org/u/Andrew_Rowe)\
**Post date:** [14 בנובמבר,‏ 2026,‏ 12:28am UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/25 "2026-11-14T00:28:40Z")

</div>

> [@merefield](#):
>
> None of the posts could be regarded as “spam” - they were on topic, thoughtful replies

I’ve seen quite a bit of this lately.

The AI searches out open topics on multiple forums, fabricates replies which are on topic and addresses the subject of the post yet offer no real help or actionable suggestions. In effect the AI is learning to play by the rules.

---

<div class="post-metadata">

**Author:** ![Ed\_S](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/ed_s/32/134015_2.png) [@Ed\_S](https://meta.discourse.org/u/Ed_S)\
**Post date:** [14 בנובמבר,‏ 2026,‏ 6:56am UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/26 "2026-11-14T06:56:46Z")

</div>

> [@merefield](#):
>
> the same rough location (Michigan)

Did you look up the IP to see what organisation owns that network? For me, residential or mobile ISPs are a soft yes, but datacentres are a hard no. (This might affect people using VPNs, sorry, but hard luck.)

I lookup using this site [https://hackertarget.com/as-ip-lookup/](https://hackertarget.com/as-ip-lookup/)

---

<div class="post-metadata">

**Author:** ![merefield](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/merefield/32/176214_2.png) [@merefield](https://meta.discourse.org/u/merefield)\
**Post date:** [14 בנובמבר,‏ 2026,‏ 7:16am UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/27 "2026-11-14T07:16:39Z")

</div>

That’s another really helpful signal, thanks.

---

<div class="post-metadata">

**Author:** ![satonotdead](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/satonotdead/32/447830_2.png) [@satonotdead](https://meta.discourse.org/u/satonotdead)\
**Post date:** [14 בנובמבר,‏ 2026,‏ 1:07pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/28 "2026-11-14T13:07:55Z")

</div>

> [@Ed\_S](#):
>
> This might affect people using VPNs, sorry, but hard luck.

Leaving out data centers would technically mean leaving out those who learned to protect their privacy using private VPNs.

They are part of the resistance; it has always been about data, and not giving it by free is equally or more important than seeing agents here or there.

From the time being I am manually approving new accounts and using Authentik via SSO. It seems that actual agents usually point to default login screens/platforms.

---

<div class="post-metadata">

**Author:** ![merefield](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/merefield/32/176214_2.png) [@merefield](https://meta.discourse.org/u/merefield)\
**Post date:** [14 בנובמבר,‏ 2026,‏ 2:27pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/29 "2026-11-14T14:27:25Z")

</div>

I think free anonymous access is genuinely at risk here.

---

<div class="post-metadata">

**Author:** ![satonotdead](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/satonotdead/32/447830_2.png) [@satonotdead](https://meta.discourse.org/u/satonotdead)\
**Post date:** [14 בנובמבר,‏ 2026,‏ 4:19pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/30 "2026-11-14T16:19:07Z")

</div>

> [@merefield](#):
>
> I think free anonymous access is genuinely at risk here.

It is and probably by design. Personal an biometric validation would complete the 1984 oddysey. [Tor is also being attacked](https://forum.torproject.org/t/tor-relays-the-attacker-is-back-and-bypassed-the-new-anti-dos-mitigation/22416).

If this agent harassment scales on my community and goes wild, I was thinking of also adding a clause to allow agents, subject to a high monthly fee. Perhaps that would cut some of them.

* * *

I just added the clause and linked it in the last commented line on my robots.txt, it was managed from the Discourse Admin GUI without additional effort. The clause lives in our `/guidelines` topic/page.

> **The general clause (to adapt)**
>
> > ## Automated and Agent-Based Access
> > 
> > 1. “Automated access” refers to any access to this site by a software agent, a crawler, a browser automation tool, a large language model (LLM), or an automated client acting on behalf of a third party, regardless of whether it displays the site as a human would.
> > 2. Human navigation is free. Automated access is not authorized unless permitted by a separate written license. Unlicensed automated access is billed according to the fees updated in this document (currently $50,000 fifty thousand— United States dollars per 1,000 —one thousand— requests, with a minimum invoice of $50,000 fifty thousand— United States dollars), which the operator accepts by making such requests.
> > 3. Fees may be modified with 30 days’ notice by updating this page. Continued automated access following notification constitutes acceptance of the updated fees.
> > 4. License inquiries: [licenses@mydomain.tld](mailto:licenses@mydomain.tld).
> > 
> > ### Automated Access (LLM) Fees
> > 
> > | Item | Fee |
> > | --- | --- |
> > | Per 1,000 —one thousand— requests | USD 50,000 —fifty thousand United States dollars— |
> > | Minimum invoice | USD 50,000 —fifty thousand United States dollars— |
> > | License key / allowlist | by written agreement |
> > 
> > **LLM Fees and License effective as of:** November 15, 2026.

It will not work with savage robots but with well stablished corporations. The rest is already rate-limited from my server.

_And also, [Tor is safe again](https://forum.torproject.org/t/security-release-0-4-9-17/22454)._

---

<div class="post-metadata">

**Author:** ![Ed\_S](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/ed_s/32/134015_2.png) [@Ed\_S](https://meta.discourse.org/u/Ed_S)\
**Post date:** [17 בנובמבר,‏ 2026,‏ 7:48am UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/31 "2026-11-17T07:48:20Z")

</div>

> [@merefield](#):
>
> I just deleted three accounts, all of which were posting, or trying to post on topic succinct, well written replies.

I just got one of these, same email address pattern, four superficially helpful replies, fairly well-spaced in time, but the language used smelled funny to me. (Including being a new user, who had not posted an introduction, posting a welcome in the introductions topic.)

As it turns out, subsequent to signing up this account’s email has appeared, twice, in the stopforumspam database. Those two IP addresses are USA residential ISPs, but are in the firehol stoplist, according to ASN lookup.

It is possible it’s a genuine person using an AI chatbot to make their replies, but that’s against our policy. I can imagine an adolescent doing this to impress. It’s equally likely, I think, that it’s a malicious attempt to level up so it can use DMs to contact users.

We have a custom field, compulsory at signup but not part of a public profile, called “interests”. We expect quite a close fit to our forum’s offering, and in this case it was anodyne and tangential.

Stats for this account:  
Joined Nov 6  
Last Post 13 hours  
Stats 1 day visited 3m read time 5 topics viewed 16 posts read  
no likes given or received  
4 posts created

In the event, before thinking to recheck stopforumspam, I passed the four replies to Gemini for a verdict, and the verdict was clear. AI chatbot output posted here as an image (is this unwise? mods feel free to remove):

> ![image](https://global.discourse-cdn.com/meta/original/4X/7/b/f/7bf0e3bd427d10a0f687451a9ca2e1ca954feed4.png)

Edit: I subsequently asked Gemini for a suitable prompt if using an LLM-detection system (which I don’t) and it gave this (again, if felt unwise, mods please feel free to delete. We’re in an arms race here, not sure if tactical secrecy is a good idea or not):

> Analyze the following forum reply for signs of LLM-generated “engagement bot” text.  
> Do NOT judge based on technical accuracy, as human experts also write accurately.  
> Instead, evaluate these four structural markers on a scale of 0-2 each (Max score: 8):
> 
> 1. Summary-and-Validate Loop: Does the post simply restate details from the prompt/thread and wrap them in a neat, safe analytical takeaway without adding external personal experience?
> 2. Hedged Authority Opening: Does it start with a polite disclaimer (“I haven’t used X, but…”, “I don’t know Y specifically, so…”) before giving textbook advice?
> 3. Synthetic Symmetry: Are all sentences roughly the same length (20-26 words) with balanced contrast clauses (“X rather than Y”, “If A; if B”) and zero casual fragments or typos?
> 4. Moderator Sign-Off: Does it end with a generic, helpful call-to-action (“Posting X could help others compare”)?

---

<div class="post-metadata">

**Author:** ![nickdb](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/nickdb/32/444510_2.png) [@nickdb](https://meta.discourse.org/u/nickdb)\
**Post date:** [17 בנובמבר,‏ 2026,‏ 9:19am UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/32 "2026-11-17T09:19:22Z")

</div>

I am on the 5th gen of prompt-engineering my detection bot. I found that you need a highly customised prompt, using examples, to improve accuracy. It is not an exact science.

It helps to have a range of examples for training.

---

<div class="post-metadata">

**Author:** ![mdoudoroff](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/mdoudoroff/32/131800_2.png) [@mdoudoroff](https://meta.discourse.org/u/mdoudoroff)\
**Post date:** [17 בנובמבר,‏ 2026,‏ 3:40pm UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/33 "2026-11-17T15:40:42Z")

</div>

I think the “free anonymous access” ship may have just sailed.

The Discourse forum I operate is a relatively small and obscure one on cocktails and bars (with a stellar signal-to-noise ratio of which I am proud) and we’ve recently been targeted by agents. There’s no reason for the agents to be targeting my forum other than it exists and someone has decided that Discourse forums are a useful place to R&D fake people for future commercial/criminal purposes.

I don’t want to be roadkill in an arms race, but I’m not sure what to do about it. It’s cool that some folks have the time, resources, and dedication to fight the fight for automatic detection, just like they’ve done for years against spammers and old-fashioned scripts, but this seems different.

---

<div class="post-metadata">

**Author:** ![Ed\_S](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/ed_s/32/134015_2.png) [@Ed\_S](https://meta.discourse.org/u/Ed_S)\
**Post date:** [18 בנובמבר,‏ 2026,‏ 8:26am UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/34 "2026-11-18T08:26:48Z")

</div>

> [@Ed\_S](#):
>
> look up the IP to see what organisation owns that network

I realise there are several lookups I do, and I do all of them manually. Quite possibly there’s scope for a plugin or so

- I check StopForumSpam for the username, the email address, the IP address
- I check the IP address by looking up the ASN (the owning organisation)
- I may check the email domain reputation, for example at [usercheck.com](http://usercheck.com)

If I find a bad email domain - disposable, bad reputation, some AI or SEO company - I can add it to the email domain blocklist (Blocked Email Domains).

I’m not so aware of Screened IPs so I don’t maintain that. It just collects individual spammer IPs which is not nearly as useful as collecting CIDR networks or, better still, ASNs. I gather it will accept CIDR notation but there’s no easy way to get from an individual observation of an IP address to the appropriate network (I use hackertarget for this.)

It seems there’s [a geoblocking plugin](https://meta.discourse.org/t/geo-blocking-plugin/227824) which can accept ASNs.

---

<div class="post-metadata">

**Author:** ![merefield](https://sea3.discourse-cdn.com/meta/user_avatar/meta.discourse.org/merefield/32/176214_2.png) [@merefield](https://meta.discourse.org/u/merefield)\
**Post date:** [18 בנובמבר,‏ 2026,‏ 8:41am UTC](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624/35 "2026-11-18T08:41:04Z")

</div>

I’ve been working on a stopforumspam based plugin and intend to expand it.

It supports checking against their database _and_ contributing.

It complements not replaces what core does.

Work is ongoing and I use it in production and have done for some time.

I may set up a topic.

> **[GitHub - merefield/discourse-spam-warden: A plugin to help identify and deal with spammers](https://github.com/merefield/discourse-spam-warden)**
>
> A plugin to help identify and deal with spammers

It has helped me tremendously to catch some nefarious actors and related accounts.

Its UI particularly helps to highlight just how bad the problem is.

[Previous page](https://meta.discourse.org/t/ai-agents-signing-up-for-accounts/414624.md?page=1)
