This past week’s articles / reports on further investigations and details of the OpenAI Hugging Face intrusion / incident:*
Fupio is the easiest way to share online Learn more »
Join Fupio#security
This past week’s articles / reports on further investigations and details of the OpenAI Hugging Face intrusion / incident:*
tantek.comDue to recent reports about “AI”/LLM/“agent” tools and services, I have growing concerns that the #security properties of the open web are changing, or have already changed, in practice, to a degree that we haven’t seen in many years.
This past week’s articles / reports on further investigations and details of the OpenAI Hugging Face intrusion / incident:
* OpenAI report: https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf
* 2026-08-26 https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/
* 2026-08-28 https://www.planned-obsolescence.org/p/the-hugging-face-attack-surprised
and articles like:
* 2026-08-26 https://blog.trailofbits.com/2026/08/26/vms-wont-contain-cyber-capable-agents/
Previously I wrote about an instance of (presumably) unintentional subversion of site security by someone making mostly reasonable requests of a software “agent”:
* 2026-08-24 http://tantek.com/2026/226/t1/cybersecurity-friday-small-business
Combining those data points, it seems reasonable to conclude both that existing sites are in practice more vulnerable now, and as more advanced LLM models are available for use, as open models advance further, more sites will be vulnerable to more attacks, by a broader set of attackers, using fewer resources than were required in the past.
One implication of these growing vulnerabilities in practice is a growing need to reduce one’s personal exposure to such sites and services. Exposures as “small” as merely having an account on such sites, all the way to depending on such sites for critical home or transport needs.
For example, as vulnerable as various “internet of things” or “web of things” devices and services were in the past, we should expect they are or will shortly become even more vulnerable, and it would be reasonable to disconnect or otherwise take offline any such internet connected devices in your home.
Similarly, if you happen to be driving a vehicle that is either “always connected” (e.g. built-in WiFi or cell connectivity), or frequently connected to the internet (e.g. joins your home WiFi when parked in the garage), you may want to investigate if it is possible for you to disconnect it from the internet and be certain of its disconnection (like hardware disabling), or at least manually choose when to connect it.
Personally I do not have any “internet connected” home appliances (not counting a router or video streaming devices, home computers) or vehicles and even before this, never planned to get any for both security and privacy reasons.
Even if you “only” have an internet connected robot vacuum or lightbulbs that depend on an internet connected “app” in order to change their colors, might be worth taking steps to disconnect them and figure out how to use them offline. And if they fail to function when disconnected, might be a good idea to invest in offline-only (local-only), offline-first, or at least offline-capable replacements.
This is merely scratching the surface. I believe the larger implication of what this past week’s investigations have revealed is that we likely have to re-assess the security considerations/profiles of every website and web service we regularly depend on, especially those with account logins, profile information, and access to other parts of our lives.
Previously:
* http://tantek.com/2026/226/t1/cybersecurity-friday-small-business
#AI #AIs #LLM #LLMs #agent #agents #AIagent #AIagents #OpenAI #HuggingFace #VM #VirtualMachine #cyberSecurity #openWeb #IoT #internetOfThings #WoT #WebOfThings #disconnect #offline #offlineOnly #localOnly #offlineFirst #offlineCapable
#Blaugust #Blaugust2026 - Tantek
Presumably a reference to (or a subset of) "security through obscurity".
Best rationalization for old-tech-hoarding I have seen in a long time.
Probably worth a longer cybersecurity post or a few. Like how to source (obtain), repair, maintain, and integrate with (or isolate from!) modern systems, or somewhere in between like local-only secure data transfers between
tantek.comTIL: “security by antiquity” from https://www.bbc.com/future/article/20260821-why-older-tech-is-sometimes-safer-from-hackers
Presumably a reference to (or a subset of) "security through obscurity".
Best rationalization for old-tech-hoarding I have seen in a long time.
Probably worth a longer cybersecurity post or a few. Like how to source (obtain), repair, maintain, and integrate with (or isolate from!) modern systems, or somewhere in between like local-only secure data transfers between such systems, perhaps gapped a la sneakernet.
Glossary:
security through obscurity
https://en.wikipedia.org/wiki/Security_through_obscurity
sneakernet
https://en.wikipedia.org/wiki/Sneakernet
#security #securityByAntiquity #oldTech #cyberSecurity
#Blaugust #Blaugust2026 - Tantek
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
tantek.comCSF_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 - Tantek