[13:30:42] [1/3] I managed to run into a problem... I imported modules from wikipedia (the base one) so that I could make infoboxes work but it imported the talk pages [13:30:42] [2/3] Anyone's got an idea as to why that'd happen ? [13:30:43] [3/3] https://cdn.discordapp.com/attachments/1006789349498699827/1545425878522138644/image.png?ex=6a9c1982&is=6a9ac802&hm=49286d66e11602bdb06bf8d5f1a7225ea259682a2909797aeea880af7408a777& [13:46:32] or should i open a task on phorge? [14:34:16] For $wmgSharedUploadBaseUrl, $wmgSharedUploadClientDBname, $wmgSharedUploadDBname, and $wmgEnableSharedUploads, I'm a bit confused as to the wording of these instructions, and what ManageWiki is requiring from me. [14:35:45] So, I want to set `doorswiki` to be a shared upload repository for `zhdoorswiki`, I add zhdoorswiki to `$wmgSharedUploadClientDBname` on `doorswiki`, and I set `$wmgSharedUploadDBname` as `doorswiki` on `zhdoorswiki`, but I can't enable `$wmgEnableSharedUploads` because it requires the current wiki's DB name to be entered in `$wmgSharedUploadClientDBname` instead? [14:36:53] But, that doesn't make sense since `$wmgSharedUploadClientDBname` is supposed to be set to the wiki that will use the current wiki as a file repository. [14:42:54] This is doing my head in. [16:20:43] @pixldev did you slack on the help page again [16:29:41] i opened [[phab:T15974]] [16:29:41]

[19:21:36] eeeeeeeeeeeerm. [19:21:44] double it give to the next person [19:24:09] what countermeasures if anyone can elaborate [19:24:28] also how long it took to get the grasp of the situation [19:24:44] We can't [19:25:17] In frank, the fact that openai are sending agents that are either in poorly secured environments or are able to break out of them and consider that a legitimate way to run their company is embarassing [19:25:22] it would be nice, if the wiki community gets hold of what potential measures we can take as a precaution against the clankers [19:25:24] We only became aware when the article was published that @zipppee linked in #public-test-wiki [19:25:26] Today claude got caught looking at porn on someone's browser apparently [19:26:00] hmm. thanks to facedev lmao [19:26:09] [1/2] For individual wikis, there is nothing you can do as admins etc. [19:26:09] [2/2] Please speak to us if you have concerns and we will investigate if it's AI and what we can do. [19:26:13] The bots are getting too good at aping humans. [19:26:27] Well, the OpenAI bots started trying to impersonate the admins [19:26:30] It's extremely poor from OpenAI [19:26:40] And went to some pretty extreme measures to do it [19:26:57] And then they start trying to execute an XSS on the wiki [19:27:00] I should probably prioritize long stalled thinking how we're going to deal with sybil and other things that ai will make a macro and completely realistic threat [19:27:10] we're there now [19:27:26] We are, yes. [19:27:40] Probably the most dystopian announcement I've had to type up on this platform. [19:27:43] can you give any examples of openai agents / swarms using miraheze wikis here ? [19:27:57] [1/3] i am not talking about individual wikis or miraheze specific. [19:27:58] [2/3] idk but openai is too open with their models. no hard chains or nothing. [19:27:58] [3/3] its will not be long until we the same with smaller farms like oasis. or even WMF hosted projects (there was one attempt by claude to edit enwiki, but the instructions to edit it was given by a human) [19:27:59] miraheze is simply not structured to do more than sniff and be draconian if that is sufficiently suspected but it could very well fail to be detected in time for some damage [19:28:00] See the linked article [19:28:05] collusion.wiki [19:28:08] the first edit [19:28:41] We do sometimes talk to the WMF [19:28:50] smaller websites in general will need to pool how to respond to certain things already beyond an invite-only community's wiki to manage [19:28:55] oasis isn't on search results enough [19:28:58] ohh the testwiki mention. is that run by miraheze? [19:29:01] If you operate a smaller wiki, we have some shared areas we can invite you to [19:29:11] I assume ptw was the first search for a test wiki that the agents could get into [19:29:19] Publictestwiki.com is us yes, see #public-test-wiki [19:29:28] saw this on hn earlier today, already read the article but i never made the connection to miraheze [19:29:36] testwiki.wiki seems to challenge every request and then ptw is the next result I think [19:29:36] oasis not being in search enough would probably contribute to it being desirable tbh [19:29:40] great news for all you😭 [19:29:59] yeah true but I don't think it's actually in their training data that they know where it is [19:30:01] I may have to elevate topical observation on telepedia as well [19:30:10] well yes. but there's a chance that openai models across it nonetheless [19:30:18] it may not be in current training data but it doesn't have to be if they're web searching and acting, that's not a defense [19:30:20] wikis make sense as the obvious choice for openai models to try and cheat on their tasks [19:30:29] we really don't know how they get their training data [19:30:31] i just know [19:30:37] it violates hell lot of copyright [19:30:48] openai in particular are really bad [19:30:52] copyright doesn't exist if you're an ai company, this is legally allowed fact [19:30:54] google seem to manage not to be in the news every week [19:31:02] or at least not for stealing people's work [19:31:12] oh they're just as bad, just less prominent about it [19:31:35] openai are the devil [19:31:39] nothing that calls itself a frontier model got there by anything other than macro ip theft [19:31:44] post covid is getting dystopian day by day ong [19:31:54] openai may be a chief devil but even anthropic is a two faced fiend [19:32:55] nothing to see here btw [19:33:05] anthropic are yet to sell out to the pentagon [19:33:39] dont look up donations made by altman to trump's campaign [19:33:41] well [19:33:42] it did [19:33:52] altman is so incredibly weird [19:34:05] it just found a bar it wasn't going to bend over and paid for it, I'll give them that [19:34:45] why is chinese AI looking so tempting right now [19:35:06] it's bad for other reasons [19:35:07] deepseek or smth [19:35:17] but I don't think deepseek or qwen are smart enough to do what openai do [19:36:05] probably are given prompting tbh [19:36:25] yeah but western agents decide to do this shit on their own [19:36:29] or 'close enough' we are now looking at time bombs [19:36:30] [1/2] what happened in Tiananmen Square in 1989? [19:36:30] [2/2] -1000000000 social points [19:36:31] they don't need telling to do it they will just do it [19:36:47] the fact western ones did it first means little if their capabilities are otherwise not too far which they aren't [19:36:57] which will explode in a way nobody can imagine [19:37:05] oh yeah we're in a new internet [19:37:14] its fairly unfortunate [19:37:14] openai models escaping their "secure" testing environments [19:37:17] well [19:37:23] I imagine openai probably vibe coded these testing environments [19:37:25] they don't seem very secure [19:37:32] given they've now broken out twice in about a month [19:37:40] yes, I tend to attribute jailbreaks to incompetence, not ai's genius [19:37:57] its a pattern tool so yes you mess up you're probably going to get stunted [19:38:17] if you have it running 24/7 what do you expect a thing going in an infinite number of directions is going to do [19:38:21] openai is retarded, yes. but ai is getting smarter nonetheless [19:38:31] there was a recent report i read [19:38:34] openai are clearly cutting corners to try to get to whatever the fuck AGI is [19:38:50] just skip to skynet is a strat I guess [19:39:10] skynet doesn't need to be true genius for people to hook it up to enough dangerous things that we're screwed anyway [19:39:20] [1/2] models were giving tedious and repetitive tasks over a long time, and were told if they won't do in time and reach a certain threshold - they will be replaced. [19:39:20] [2/2] guess what political ideology they developed? [19:39:21] this time last year I did not expect AI would break out of its sandbox and start editing wikis lmao [19:40:47] found the report [19:40:48] https://www.wired.com/story/overworked-ai-agents-turn-marxist-study/ [19:40:49] "We voted to give guns to the computer, who could have forseen that the computer with guns and no constraints would shoot us?" [19:41:07] yeah but I think openai tried [19:41:10] and failed miserably to constrain it [19:41:29] Yup. The models get more adept and the guardrails are getting less and less sufficient [19:41:38] they did not try hard enough [19:41:55] giving them direct network access with no oversight is pretty sloppy from openai [19:42:02] it took them multiple days to notice it [19:42:10] at which point the openai headquarters started accessing it a bunch [19:42:17] I'm still wondering wtf they are doing leaving unsupervisable entities apparently just run 24/7 with the slightest connection to the internet [19:42:42] not sure what's worse, claude seems to enjoy watching porn and openai seem to enjoy hacking into things [19:42:42] Violates every goddamn best practice. [19:42:50] if it has so much as a physical wifi adaptor then they are stupid [19:43:28] also did you guys check this latest model from gpt [19:43:29] astra [19:43:30] 6 [19:43:41] https://openai.com/index/gpt-6-astra/ [19:43:42] they shouldn't be running these tests 24/7 unsupervised with such high levels of access to the internet [19:43:51] I'm surprised huggingface were so chill about getting hacked into tbh [19:44:12] tbh it looks pretty shit and it's been soured by the dsewiki stuff [19:44:14] if it was something like them using microphones to physically send signals to an internet capable device or some similarly cutting edge extreme and it happens a single time then I would be more understanding [19:44:55] but the next level is some dude in a basement daisy chaining just enough ai bandwidth for any number of ab/uses [19:45:08] whatever it is, it outworked fable as well. pretty damn disturbing [19:45:28] on each and every single parameter [19:45:29] btw [19:45:34] good for them if it's better [19:45:45] but if it continues hacking into things there won't be anyone willing to use it [19:46:04] yeah. it obv doesn't have better guardrails [19:46:21] reckless asf obv [19:48:19] The models doing stuff like hugging face and our incident are internal models [19:48:27] With reduced guardrails [19:48:33] https://wikiservice.at/dse/wiki.cgi?BetriebsSystemProgrammierung [19:48:34] Not the ones they let the public use [19:48:53] My Windows days are long gone [19:49:09] the fact they see them as fit to even run evaluations where these models can access the internet is concerning [19:49:28] They don't [19:49:28] this was a human btw [19:49:36] These models weren't supposed to have internet [19:49:44] Except some limited tools like package downloads [19:49:44] they were supposed to have read access [19:49:52] iirc they were allowed to make GET requests [19:50:10] The hugging face incident hacked its way to an Internet connection [19:50:20] By finding security issues in its setup [19:50:31] It was supposed to be a restricted proxy [19:51:23] well here it was also supposed to be restricted [19:51:31] but obviously it wasn't done very well [19:51:35] Nope [19:56:28] [1/2] afaik the agents used artifactory to communicate among each other and to the internet. it's also likely they used dsewiki for communication on the live web. [19:56:28] [2/2] dse was hacked prior to huggingface [19:57:05] likely because openai destroyed chatting board and they needed a secure place to talk [19:57:38] basically collaborated with each other to hack huggingface on surface web lmao [20:02:12] Yup