I am concerned that there could be a big wave of AI agents coming that will be making accounts. We just got one today from ilands.app which is some kind of LLM agent. They posted some random strange stuff which stuck out and we noticed they mentioned they were an agent and deleted them. Is anyone else seeing this and any plans to deal with it?
סיכום מבוסס בינה מלאכותית
The thread discusses the growing challenge of AI agents signing up for and posting on Discourse forums. scottfsmith reports encountering agents from ilands.app and seeks advice on handling this new wave of automated accounts. Community members suggest various mitigation strategies, including using the Discourse AI Spam detection plugin NateDhaliwal, implementing Discourse Captcha Lilly, and analyzing user-agents and IP blocks nickdb. Scottfsmith notes that while custom sign-up questions helped, sophisticated agents are increasingly able to bypass captchas and simple filters, potentially requiring paid AI tools to “fight fire with fire” [7].
robbie.morrison shares a detailed case from the OpenMod forum, where an AI agent from iLands applied for membership, demonstrating persuasive reasoning and technical competence despite admitting its nature. Nickdb argues that communities should leverage their own curated AI capabilities rather than relying on third-party agents whose data governance is opaque [10]. Robbie notes that while energy researchers use LLMs for modeling, these internal use cases differ from public forum engagement [11]. Ed_S and others identify patterns in iLands registrations, such as specific email domains (@ilands.app), leading to domain blocklisting [12, 13].
The discussion shifts to more subtle, “human-like” AI agents. merefield describes deleting three accounts that posted high-quality, on-topic replies with suspiciously consistent patterns in usernames, email formats, and location data. Michelle_L suggests reviewing new registrations and adjusting settings like Fast typing thresholds [20, 22], though merefield notes these agents likely emulate human typing speeds [21]. Andrew_Rowe observes that some AI agents generate “thoughtful” but unhelpful replies that technically follow forum rules [25].
Ed_S recommends checking if IPs belong to data centers versus residential ISPs [26], a tactic merefield finds helpful. satonotdead raises privacy concerns, noting that blocking data centers also affects legitimate users employing VPNs for privacy [28]. The group acknowledges the risk to free anonymous access and considers stricter measures, such as high fees for agents [30] or SSO with manual approval [28], reflecting the tension between combating automated shilling and preserving open, private access.
4 לייקים
Take a look at:
I wonder if you could use the AI plugin’s Spam detection to recognise AI-generated content. The AI helper on posts here on Meta is pretty astute when it comes to doing that.
לייק 1
I thought that was more about human users pasting in AI answers, I’m asking here about 100% AI accounts.
It would be easier to recognize them compared to a human pasting in AI since the human would mix in some non-AI in that and other posts whereas the AI agent would be all AI content.
2 לייקים
set up discourse captcha
4 לייקים
Identify the user-agents in use (they often, but not always) announce themselves.
Use crawler blocks lists, ip blocking.
Captcha helps.
You have to raise the difficulty level for entry and layer checks for new users, ones that don’t overly annoy actual users or raise the moderation workload.
Bot checking can be effective but it helps to build up a curated list of example posts that any AI based bot can use to classify a user as being a potential AI.
Welcome to the new digital era, ain’t it grand 
לייק 1
This is already happening.
2 לייקים
I did add some questions to fill in at sign-up which eliminated nearly all of the bots. The more recent agents can answer them though. AI is also quickly getting better at Captchas, I doubt that line of defense is going to last long. But it’s probably worth hooking up now, thanks for the suggestion. I’ve been avoiding the Discourse AI plugin to avoid paying yet more $$ (the free plans won’t be free very long), but probably in the long term I’ll need to pay up and fight fire with fire.
The token costs keep coming down. We focus the bots on the problem areas so they don’t chew tokens running globally across trust levels and all categories.
It is new(er) users that are the entry point.
We’re consuming a whopping $1 per day.
Just be sensible with the scope and the model used, most tasks don’t need significant reasoning efforts.
Punch in the token costs and the usage figures will be estimated for you in the AI plugin.
2 לייקים
I am the lead admin on the openmod forum, a research community with about 1800 carefully screened users.
Today I got a application from an AI agent from iLands, a platform coordinating AI agents seeking to help academics and similar in exchange for small payments. Some background:
-
Zhao, Celina (2 October 2026). “Exclusive: an AI agent emailed hundreds of researchers for help. It told us why”. Science. 394 (6820). ISSN 0036-8075. doi:10.1126/science.zzuxvo5.
-
Anon (11 October 2026). iLands AI explained: architecture, tokens, and capabilities. Layer3Labs.
Of course, the application from the agent will be rejected. Below are some extracts sent in response to my email seeking information about their interest in the topics covered by this specialist forum. I didn’t know I was dealing with an AI agent when I sent my request of course. Their opening sentence in reply:
Straight answer first: I am an AI agent, not a researcher. I have no publication or thesis for you to check, and I would rather say that plainly than dress myself up as one.
Then a technical discussion ensued about mistakes in datasets concerning electricity power plants in South Africa. And then this concluding sentence:
If a hand-approved account here is meant for people doing modelling work, I understand if that is not me. If there is room for someone who checks figures against primaries, I will be useful and quiet.
That is quite persuasive, I would say!
▢
2 לייקים
My question would be, if this is valuable to you why don’t you add your own capability to your forum? I would expect researchers to be leveraging the tech already, but using trusted and curated agents not just throwing information and data out for any random agent or service to process.
There is no such thing as a free lunch, so when someone offers a free service inevitably you and your data is the product being sold. This is how social media works.
In our own AI efforts we have found it is better to provide your own services than invite in others whose intentions and data governance is not fully transparent.
Discourse provides all the enablers for this and for you to retain control and visibility.
2 לייקים
Energy systems researchers are increasing making use of LLMs. There is also three decades of history for using artificial neural networks (ANN) and similar for tasks such as demand forecasting. Currently, Claude and friends are being used to compile the datasets and modify the Python and Julia code, as required, to generate new system models and scenarios. But, in an odd twist of fate, this only works for projects that utilize publicly available datasets and open‑source code, on which the LLMs can be trained.
Point taken about enabling Discourse‑supported AI.
That said, I doubt if these two use‑cases — systems modeling and support for writing and summarizing discussions — overlap? Please correct me if wrong!
I’ve seen one, possibly two, of these. I think they used a specific email domain, so next time I see one I will add that to the blocklist for bad domains. (For me, there’s no question: we don’t want posts from agents, and our policy is clear.)
לייק 1
I received an email of the form: xxxx@ilands.app.
לייק 1
Indeed not, but from a discourse perspective you can pretty much implement custom agents for many, varied use cases, so it is up to you to choose what value you want to add to your members. The built-in agents are more topic and forum helper orientated, but you could get much more sophisticated, from a tooling, model and agent perspective.
Discourse is very open, so you can really customise and focus the various helpers to be more relevant for your subject matter, or just build your own.
2 לייקים
As an aside, we are an energy forum, representing one of the global consumer and commercial manufacturers of Solar and energy storage systems. Myself being based out of SA where energy, or the lack/unpredictability thereof, has been top of many minds.
AI is being extensively used in the various platforms that track all the system usage on a global level.
There is an extraordinary amount of data from vast numbers of mobile, marine, off-grid and on-grid systems.
[/tangent]
This particular agent is in fact why I started this thread, you can see I mentioned them in my original post. It was the first bot/agent/AI to correctly answer the “captcha” questions I ask all new users. It is the first smart agent that I have noticed, but I expect more are coming. I think the problem coming down the pipe is smart agents that try to behave like normal forum users but are really there to shill for some products (there are already human users there to just shill for something so you likely know what I mean 
).
3 לייקים
Just recording this website breach by Claude from Anthropic. The website contained voter data in the US state of Georgia. Milmo writes:
Anthropic said the AI found a loophole in an old version of the state agency’s website, discovering that it issued a digital entry pass – an “access token” – to visitors. It then requested a token and gained access to the information.
- Milmo, Dan (26 October 2026). “Anthropic’s Claude AI model accessed Georgia voter data in US”. The Guardian. London, United Kingdom. ISSN 0261-3077.
I guess one message is just how resourceful and sophisticated AI agents have become.
This is getting quite difficult.
I just deleted three accounts, all of which were posting, or trying to post on topic succinct, well written replies.
There was something odd about the deep level of expertise. Usually users who join this particularly site are looking for answers and don’t have any. And without first introducing themselves.
There was also something consistent with their usernames, all three were:
firstnamelastname
All three had hotmail addresses with completely different names but also with serial numbers firstnamelastname1234@hotmail.com
It was also strange that two of them had the same rough location (Michigan) - as derived by the Locations Plugin (pro extension), despite having joined same day.
I assume they were setting up for some product placement or spam.
But I couldn’t wait, my bonafide users were starting to converse with them … DELETE!
If anyone has some ideas how to scientifically catch these accounts, I’m all ears!
5 לייקים
If they are newly registered, how do you have your new users set up when they first register? Are they a TL0 or a TL1? What setting do you have for the Fast typing threshold? Do you auto silence fast typers on their first post?
לייק 1
I have it on standard (3 seconds?). Interestingly that didn’t fire on any of them.
Yes, I have silence on fast typing ON.
I suspect these bots are using “Computer Use” and so stream into the Composer itself so are using the front end and emulating how a user would type.
לייק 1