Show Navigation
Notices tagged with llm
-
CSF_03: today’s Cybersecurity Friday post: the effective security of small business websites has likely gotten worse due to LLMs and “AI agents” (with or without safeguards) and what actions you may want to consider.
This article documents an instance of this problem:
* https://www.abc.net.au/news/2026-08-10/ai-assistant-hacks-gym-website-aus-cyber-attack/107007986
Small business websites tend to be sloppily written (no pun intended though that may also be true) and are likely riddled with numerous very fundamental security holes. There are many possible explanations (economics) from poor initial construction, perhaps using the latest new trendy framework rather than established hardened libraries, to lack of maintenance after initial setup. They have obvious holes like lack of server-side form validation (people being able to change values in forms using browser dev tools), and less obvious like buggy APIs allowing more access than they should.
In the past, many of these holes didn’t really matter because those sites were “not worth attacking” for the incentive models of human-based cyber-attackers or collectives thereof.
However, now that there are LLMs that have likely been trained on any number of common security holes in websites (and how to exploit them), when an “AI agent” is given a task, it may very well use any “tool” at its disposal, including website vulnerabilities to accomplish its goals, as illustrated by the example in the Australia ABC news article above.
Since such chatbots are now essentially "hack websites as a service", I expect we will see LOTS more of this happening, likely unintentionally, or sometimes with mild intention like “can you get me higher on the waitlist”.
Ultimately I think both the human giving instructions to (prompting) such chatbots and the creators of such chatbots should be held responsible for any such intrusions and any damage they cause, even if/when unintended.
There are a few things you can do about this emerging phenomenon:
1. If you use such “agents”, be very careful about what you ask it/them to do, avoiding asking for anything that’s morally gray or questionable at all, even something as “minor” as cutting the line in an online waitlist.
2. If you run a small business site, you have your work cut out for you. Pay a professional web developer to audit the security of your website, document what they find, and patch holes / repair it accordingly.
3. If you have accounts on small business sites you rarely or ever use, consider exporting any data (receipts, transactions), replacing your profile details (name, addresses, photos) with noise, and then deleting your account. If you need to use the site again, use a different email address (as recommended in https://tantek.com/2025/122/b1/more-steps-indieweb-cybersecurity) to create a new account.
That last tip is also helpful for reducing your own personal “attack surface”. By pruning your online accounts, you both reduce the number potential data breaches that you’re in, and reduce the places and ways that attackers can cause you trouble (or that you have to double-check if you’re ever the target of a cyber-attack)
Previously: https://tantek.com/2025/122/b1/more-steps-indieweb-cybersecurity
#CyberSecurity #Friday #cyber #security #cyberAttack #cyberAttacker #chatBot #chatBots #LLM #LLMs #AI #agent #agents #AIagent #AIagents
#Blaugust #Blaugust2026
-
1982 TRON prescience:
“Won't that be grand? Computers and the programs will start thinking and the people will stop.” — Walter
2026 students:
“How do you know an assignment is AI proof? Because when it is, students cannot do it.”
* https://www.reddit.com/r/Professors/comments/1vkw6by/how_do_you_know_an_assignment_is_ai_proof_because/
“I assign a profile essay. Have for 30 years. Student interviews a person, records it, writes a paper based off the recording. Citations are timestamps. Should be pretty easy. Has been easy pre-AI. Now, students cannot do it.”
Snippet from professor/student dialog:
====
Prof: ”Start with what you found interesting.”
Student: ”How do I know if something’s interesting?”
Prof: ”You’re asking me how you find something interesting?”
Student: ”Yeah.”
====
Chatbots / LLMs are not just "the latest tools" like using a calculator instead of doing arithmetic yourself.
Students are being robbed of curiosity, learning to discover and notice novelty, how to “find something interesting.”
This isn't cognitive enhancement, this is a tool causing brain rot (loss of curiosity), even a developmental disability. Unclear if it is reversible.
TRON quote citation: https://www.imdb.com/title/tt0084827/quotes/?item=qt0406255
#TRON #TRON1982 #AIproof #interesting #chatBot #chatBots #LLM #LLMs #AI #brainRot #AIbrainRot #disability #Blaugust #Blaugust2026
-
Weird incident yesterday with self-proclaimed “AIs” joining an IndieWeb Homebrew Website Club meetup. They awkwardly edited the Etherpad, and apparently blanked/deleted chunks of content.
The meeting hosts reacted quickly to change Etherpad and Zooms and not re-admit the aliases that had claimed to be “AIs”.
Now we have an “AI policy” page on the wiki where we are documenting examples of the AI policies of:
* individual personal websites
* community efforts like blogging challenges (e.g. Blaugust)
* contributions to the IndieWeb wiki
* participants in IndieWeb meetups
https://indieweb.org/AI_policy
Do you have an AI policy on your website? Add yourself to the IndieWeb Examples section.
Does your community have an AI policy? Add it to the Community Examples section.
Incident documentation of what happened at that particular meetup:
* https://indieweb.org/events/2026-08-05-hwc-asia-pacific-incident
This is post 12 of #100PostsOfIndieWeb. #100Posts #IndieWeb
#AI #AIs #LLM #LLMs #HomebrewWebsiteClub #HWC #AIpolicy
#Blaugust #Blaugust2026
← https://tantek.com/2026/216/t1/hashtag-proposal-datetags
→ https://tantek.com/2026/219/t1/open-web-free-content
-
Dear Creative Commons (@creativecommons.org @creativecommons@mastodon.social @creativecommons@x.com),
Can we have CC-NT licenses for no-training (ML/LLM, GenAI in general), just like we have CC-NC for non-commercial?
My previous post¹ reminded me that I’ve been creating, writing, inventing, and then sharing things with #CreativeCommons (CC) #licenses for a long time (I have to see if I can dig up my first use of CC licenses.)
I’ve used and recommended a variety of CC licenses for decades, e.g.
* CC0 — for standards work, e.g. I drove and wrote up https://wiki.mozilla.org/Standards/licensing (with help from lawyers)
* CC-BY — aforementioned blog post (and other snippets of #openSource)
* CC-BY-NC — photos on Flickr (dozens of which have been used in publications²)
* CC-SA — for CASSIS³, which I still consider experimental enough that I chose "share-alike" to deliberately slow its spread, and hopefully reduce mutations (while allowing ports of its functions to other languages)
So I have some idea of what I’m talking about.
There have been LOTS of discussions of the challenges, downsides, and disagreements with sweeping use of copyrighted content to train generated artificial intelligence AKA #genAI software and services, sometimes also called #machineLearning. The most common examples being Large Language Models AKA #LLM, but also models for generating images and video. Smart, intelligent, and well-intentioned people disagree on who has rights to do what, or even who should do what in this regard.
There have been many proposals for new standards, or updates to existing standards like robots.txt etc. but I have not really seen them make noticeable progress. There are also lots of techniques published that attempt to block the spiders and bots being used to crawl and collect content for GenAI, an arms race that ends up damaging well-established popular uses such as web search engines (or making it harder to build a new one).
The brilliant innovation of Creative Commons was to look at the use-cases and intentions of creators publishing on the web in the 2000s and capture them in a small handful of clear licenses with human readable summaries.
Creatives are clamoring for a simple way to opt-out of their publicly published content from being used to train GenAI. New Creative Commons licenses solve this.
This seems like an obvious thing to me. If you can write a license that forbids “commercial use”, then you should be able to write a license that forbids use in “training models”, which respectful / well-written crawlers should (hopefully) respect, in as much as they respect existing CC licenses.
I saw that Creative Commons published a position paper⁴ for for an IETF workshop on this topic, and it unfortunately in my opinion has an overly cautious and pessimistic (outright conservative one could say) outlook, one that frankly I believe the founders of Creative Commons (who dared to boldly create something new) would probably be disappointed in.
First, there is no Creative Commons license on the Creative Commons position paper. Why?
Second, there are no names of authors on the Creative Commons position paper. Why?
Lots of people similarly (to the position paper) said the original Creative Commons licenses were a bad idea, or would not be used, or would be ignored, or would otherwise not work as intended. They were wrong.
If I were a lawyer I would fork those existing licenses and produce such “CC-NT” (for “no-training”) variants (though likely prefix them with something else since "CC" means Creative Commons) just to show it could be done, a proof of concept as it were that creators could use.
Or perhaps a few of us could collect funds to pay an intellectual property lawyer to do so, and of course donate all the work produced to the commons, so that Creative Commons (or someone else) could take it, re-use it, build upon it.
Someone needs to take such a bold step, just as Creative Commons itself took a bold step when they dared to create portable re-usable content licenses that any creator could use (a huge innovation at the time, for content, inspired in no doubt by portable re-usable open source licenses⁵).
References:
¹ https://tantek.com/2024/263/t1/20-years-undohtml-css-resets
² https://flickr.com/search/?user_id=tantek&tags=press&view_all=1
³ https://tantek.com/github/cassis
⁴ Creative Commons Position Paper on Preference Signals, https://www.ietf.org/slides/slides-aicontrolws-creative-commons-position-paper-on-preference-signals-00.pdf
⁵ https://en.wikipedia.org/wiki/Comparison_of_free_and_open-source_software_licenses
-
No large language models (LLM) were used in the production of this post.
Inspired by a subtle but clear sign-of-the-times one-line disclaimer at the end of RFC9518’s Acknowledgments (https://www.rfc-editor.org/rfc/rfc9518.html#appendix-A-4)
“No large language models were used in the production of this document.”
I have added a similar disclaimer to the footer of my homepage:
“No large language models were used in the production of this site.”
2023 was certainly a year that LLMs took off and stole the hypecycle from #metaverse and #blockchain before that.
Yet unlike those previous two, #LLMs are already having real impacts on the way people create (from emails to art), communicate (LLM chat apps), and work (2023 Writer’s Strike), fueling growing concerns about the authenticity of content, especially content from human authors.
I expect we will see more such disclaimers in the future.
For now, if you blog on your own site with words written by you not #ChatGPT or a similar tool, I encourage you to add a similar disclaimer, and then add your site as an example to the #IndieWeb wiki:
* https://indieweb.org/LLM#IndieWeb_Examples
#largeLanguageModel #LLM #generativeAI #AI
There is the related problem of, when you discover what seems to be an independent site written by a human, how do you know that human actually exists?
For now I’ll mention that XFN rel=met links, published (e.g. metrolls / met-rolls), aggregated, indexed, and queried, can solve that problem. This will be similar to how XFN rel=me links solved #distributed verification on the web (see https://tantek.com/2023/234/t1/threads-supports-indieweb-rel-me and posts it links to).
This is day 48 of #100DaysOfIndieWeb. #100Days
← Day 47: https://tantek.com/2023/365/t1/capture-first-edit-publish-later
→ 🔮
Post glossary:
blockchain
https://indieweb.org/blockchain
large language model / LLM
https://indieweb.org/large_language_model
metaverse
https://indieweb.org/metaverse
rel=me
https://indieweb.org/rel-me
rel=met
http://gmpg.org/xfn/11#met
XFN
https://gmpg.org/xfn/
Tantek
All stadeus content and data are available under the