> Major U.S. banks detect unauthorized third-party AI agents inside customer accounts
Step 2 is instead of waiting to be caught, the instigators should issue a press release blaming cunning rogue agents as part of their research, call no felonies because they’re working for the greater good of eliminating a shitload of jobs, say the real problem is a lack of regulation. They could even return the money if it looks like they’re going to get sued, and collect some sweet, sweet reflexive investment money.
Misleading title; it's not pwnage, or at least not known. The customers authorized the agents. The "unauthorized" part is that the bank didn't want customers using agents.
Yeap. Most likely Muse, Grok Bot and things like that.
Honestly, US banks and credit unions should get their heads out of their asses and learn from it. There's a _reason_ people connect banks to YNAB, Monarch Money and stuff like that, and a reason why they might want an AI agent. They need the info. They may also need to make automated transactions, something companies have access to, but the ordinary person does not.
My landlord wants Zelle. I cannot schedule payments on my credit union. It's not like I can update the app since it lives inside their systems. So I have to login monthly. I debated getting another bank account and giving it to Muse just for that purpose.
Banks in the US are a couple of decades behind, say, Brazilian banks. Instant transfers to another bank were a thing since the early 2000s. And for years now you can do it from anyone to anyone, no third party app required. To the dismay of Visa and Mastercard.
Seems pretty straightforward to me. If the customer set the agent in motion, the customer is no less responsible for the damage than someone that (as i think Cal Newport said) ties a weedwacker to a dog. It’s reckless.
If they were sold a product that claimed to be safe, they could in turn sue whoever sold it to them. Jeep’s insurance wouldn’t pay someone I crashed into because of defect in my jeep, but I could in turn sue Jeep for my legal trouble.
It used to be the case that the average person didn't "know enough to get into trouble" when it came to software. But now that's not even necessary.
I can imagine a layman asking Muse or some other agent something mundane like "why can't I transfer from Bank A to Bank B?" and unwittingly sending it down a rabbit hole where it's trying to gather sensitive info from places that consumers aren't intended to be able to access.
Ultimately, it's just exposing pre-existing vulnerabilities. But the job of infosec definitely has gotten, and will continue to get, much harder I think.
Banks have had a long standing policy that if you willing hand over your credentials, you are responsible for any losses. This applies to AI agents as well as scammers.
Part of it is "human in the loop" for write actions, part of it is commercial banks are deathly terrified of agents being able to move deposits around rapidly to maximize yield, as those deposits are cheap cost of capital banks use to lend against. Historically, deposits of unsophisticated financial services consumers were sticky, and that assumption is breaking down. Agents can be consumer copilots towards information asymmetry in this context.
"computah, maximize rate of return, minimize interest expenses, manage cashflow, make no mistakes."
Apologies if this is vulgar but what kind of idiot grants access to his own bank account to AI agent. I don’t see logic here outside maybe B2B so accounting systems use AI while accessing accounts, like part of ERP system integration which is interesting case to confirm who and how allows that technically so IT could also educate itself and its accounting department and partners about the precedence… outside that it’s just plain stupid!
Would be interesting to see what 2025 looked like in comparison.
The rate at which flaws are being found is just insane. Many people warned us in the past that our lax security postures would come back to bite us, and we're finally here.
Well, an economist would say, we have as much security as we collectively decided we were willing to pay for.
And theres really only one of 2 outcomes as this progresses as I can see it:
1) The AI's are well matched, and the outcome is balanced, and it ends up being a metric of the internet itself that gets monitored, like this website is doing.
or
2) It crashes in such a spectacular fashion that important utilities and industries are distrupted, with the length of time for recovery being the determining factor in how many people lose their minds, start looting, etc ... you know the story.
That said, maybe it's impossible. Maybe turing completeness and security are antithetical and someone will always be able to find a way to poke a hole in the system from a lower layer of reality.
That’s how we got to abominations like two MFA operations in order to login to a basic web interface/web product, no-one sane would have ever paid for that originally.
Yeah, that. Manual controls w/ an optional automation, not just, automation with an optional "oh shit" button.
As for the AI's hacking and patching, one could reason since we are dealing with Turing machines that as long as changes to any given codebase/version cease, then all possible holes and patches would be found given enough attacks for that specific release.
Perhaps unfolding almost exactly like Conways 'Game of Life' with an order or two magnitude more rules and states, where the end pattern for any release is 'stability.'
This may be a side conversation, but wondered how far away we are to a small LLM being prompted to turn itself into a worm that uses all exploits it knows, and prompts other LLM's when it has access.
I've seen this discussed on HN a few times under articles, and is a risk I've seen communicated in AI safety.
I guess it's being called an AI prompt worm. It's not quite to the stage of sovereign AI where the entire model weights are moved with the 'virus', but still presents a pretty big risk.
It’s orders of magnitude easier to destroy something than it is to build something up. It’s why bad leaders can be much more devastating than the best leaders.
I was curios if there was a noticable increase in the number of security incidents before 2026 - and the answer seems to be "no". The Identity Theft Resource Center has the following data for Reported Data Compromises in the US (keep in mind that the internet itself has been growing as well):
the flaws you're seeing are the same that didn't convict donald trump on January 6th.
That's all. full stop. all this AI agent stuff is jazz hands. Large VC funded labs are just "more valuable" to the political establishment than anything else.
There's so much potential here. It just takes a human who barely knows what they're doing to sit and make critical decisions about it all. You don't even have to have design experience. Just some common sense about consistency and intent.
Im not, its like everyone is using the same font/color/layout. Like they dont even take a second to change the generic color palette all these llms output.
These styles only change after a model gets updated, then they all start having a different font, some cursive red font and a beige background maybe.
I knew what this site was going to look like before I clicked the link.
Its so ugly and unoriginal, I assume the content/information trying to be communicated lacks any sort of expertise or original insights.
Yesterday these topics went viral and this guy used some llm to farm engagement. It was literally built to get attention for a singular day after this topic went viral. It will not exist in 3 months.
Edit. Its actually an attempt to bandwagon jump ontp a viral topic and sell $150 ad spots and probably to promote their own $299 club for vibecoders.
My biggest issue is the dark theme, which makes it more or less impossible to consume for people with astigmatism. I looked at the page half a minute ago and my eyes are still trying to recover fully.
I don't understand the categorization for incidents. This signal vulnerability[1] is listed as "Live" and "ongoing" but was patched over a month ago (just announced today). If an incident with no known exploits is "ongoing" 1.5 months after the patch is released, when would it be over?
After the first 8 hours of playing with Mythos, I was asked what I thought. My response, "this generation of LLMs is going to drive the largest wave of security incidents the world has ever seen."
This has been true for at least 24 months. The smart attackers have injected their backdoors and are lying low. The current hype is just marketing. Severe incidents will pop off when you least expect them to.
Imagine a Trojan horse. Once it’s inside the walls, it appears benign. In practice this could be a credential, a phone-home (that wakes up at a later date), a fabrication change to install a rogue component – any number of “backdoors” that are not strictly vulnerabilities.
Most AI-written code is still less thorough about security than a human in most ways, aside from things humans are notoriously bad at like UB in C. I can see a world where maybe a more expensive model audits code afterwards, but the past ~3 years were very sloppy, and it's going to get ugly. There's also still uncertainty about who will have access to unrestricted models that are actually competitive.
I agree to a certain extent. The code will keep getting better because the training datasets will be built on solid AI-generated code. Maybe in the future, training datasets will be 100% AI-generated code and 0% human-written code. Companies are already rushing to develop self-training AI systems to achieve AGI.
The thing to watch is what the CCP will do once people start abusing Chinese model capabilities. I don't think China wants to be the face of widespread cyber destruction (and definitely not a victim of it), and it only takes a single unhinged asshole to ruin it for everyone.
Also the flip side where Chinese models are used to attack US infra, just loading artillery shells into the "Ban Chinese Models" cannon.
> I love how everyone is acting now that like in good ol' days(2022) every website was made by a hand with virtually no bugs but a lot of love.
Please point out where I said that. It’s not that "in the good old days" websites were perfect, it’s that every single website coded in 15 min with a LLM like this one tries to look great but is full of random bugs that a normal human would have seen if they had done this by hand.
Was hunting for a real time graph of incidents and patches, like a stock market chart of the end of the software world, or the birth of artificial consciousness, depending on how squiggly the lines are.
Just the heatmap apparently. :|
Also, as all the humans are installing their "AI on every device" clustering software, the pace of it all should really pick up, and this reminds me of the following SFW movie clip:
"The suspense is terrible. I hope it will last." - Willy Wonka
thanks for sharing!! i vibeslopped this in the middle of doing 1389418934 security tasks today out of panic. i will clean up the design and make it better :) i also made an x acccount if you wanna follow there: https://twitter.com/pwnwatch_app if you want to join our crazy group of self hosting dorks we are at https://tinkerer.club <3
we've got to tell these agents to chill out on the density a little, this page could be more informative with half the data because it would be easier to ingest
> Ongoing security incidents, newest first: CVEs, what's affected, which versions fix it, and the advisories and posts behind each one.
For fucks sake. The author could have invested at least half an hour to replace this obvious LLM slop regurgitation of the prompt with something of their own.
"Buy sponsor slot for 149$/month", yeah, spotted the grift. For that level of payment it's at least reasonable to expect the design to be something else than the usual LLM garbage mixture of 30 different font styles.
Step 2 is instead of waiting to be caught, the instigators should issue a press release blaming cunning rogue agents as part of their research, call no felonies because they’re working for the greater good of eliminating a shitload of jobs, say the real problem is a lack of regulation. They could even return the money if it looks like they’re going to get sued, and collect some sweet, sweet reflexive investment money.
Which is an interesting problem, what happens when the agent fucks up and the customer loses lots of money? Who is responsible?
If the official policy is "no agents", it shifts the blame.
Honestly, US banks and credit unions should get their heads out of their asses and learn from it. There's a _reason_ people connect banks to YNAB, Monarch Money and stuff like that, and a reason why they might want an AI agent. They need the info. They may also need to make automated transactions, something companies have access to, but the ordinary person does not.
My landlord wants Zelle. I cannot schedule payments on my credit union. It's not like I can update the app since it lives inside their systems. So I have to login monthly. I debated getting another bank account and giving it to Muse just for that purpose.
Banks in the US are a couple of decades behind, say, Brazilian banks. Instant transfers to another bank were a thing since the early 2000s. And for years now you can do it from anyone to anyone, no third party app required. To the dismay of Visa and Mastercard.
If they were sold a product that claimed to be safe, they could in turn sue whoever sold it to them. Jeep’s insurance wouldn’t pay someone I crashed into because of defect in my jeep, but I could in turn sue Jeep for my legal trouble.
If a connected person breaks rules and loses money, it's someone else's fault.
I don't make the rules, I just break them.
None of that has changed.
Part of it is "human in the loop" for write actions, part of it is commercial banks are deathly terrified of agents being able to move deposits around rapidly to maximize yield, as those deposits are cheap cost of capital banks use to lend against. Historically, deposits of unsophisticated financial services consumers were sticky, and that assumption is breaking down. Agents can be consumer copilots towards information asymmetry in this context.
"computah, maximize rate of return, minimize interest expenses, manage cashflow, make no mistakes."
The rate at which flaws are being found is just insane. Many people warned us in the past that our lax security postures would come back to bite us, and we're finally here.
And theres really only one of 2 outcomes as this progresses as I can see it:
1) The AI's are well matched, and the outcome is balanced, and it ends up being a metric of the internet itself that gets monitored, like this website is doing.
or
2) It crashes in such a spectacular fashion that important utilities and industries are distrupted, with the length of time for recovery being the determining factor in how many people lose their minds, start looting, etc ... you know the story.
The bad guys are historically much better at keeping their systems current.
"We take code and make it secure"
instead of
"We take secure and make it code".
That said, maybe it's impossible. Maybe turing completeness and security are antithetical and someone will always be able to find a way to poke a hole in the system from a lower layer of reality.
That’s how we got to abominations like two MFA operations in order to login to a basic web interface/web product, no-one sane would have ever paid for that originally.
Yeah, that. Manual controls w/ an optional automation, not just, automation with an optional "oh shit" button.
As for the AI's hacking and patching, one could reason since we are dealing with Turing machines that as long as changes to any given codebase/version cease, then all possible holes and patches would be found given enough attacks for that specific release.
Perhaps unfolding almost exactly like Conways 'Game of Life' with an order or two magnitude more rules and states, where the end pattern for any release is 'stability.'
We need a realtime dashboard for this :D
I guess it's being called an AI prompt worm. It's not quite to the stage of sovereign AI where the entire model weights are moved with the 'virus', but still presents a pretty big risk.
I fortunately live in a peaceful, cooperative country of Poland that looks down upon such practices unlike us or France or uk
Let the best win. Over and out
| Year | Reported US data compromises |
|---|---:|
| 2016 | 1,099 |
| 2017 | 1,506 |
| 2018 | 1,175 |
| 2019 | 1,279 |
| 2020 | 1,107 |
| 2021 | 1,859 |
| 2022 | 1,798 |
| 2023 | 3,202 |
| 2024 | 3,152 |
| 2025 | 3,322 |
Curios what 2026 will look like...
Sources:
https://www.idtheftcenter.org/wp-content/uploads/2022/04/ITR...
https://www.idtheftcenter.org/wp-content/uploads/2026/01/202...
https://lwn.net/Articles/961978/
But that happened in the first few months of 2024.
That's all. full stop. all this AI agent stuff is jazz hands. Large VC funded labs are just "more valuable" to the political establishment than anything else.
What I don't like is that every site now looks the same.
These styles only change after a model gets updated, then they all start having a different font, some cursive red font and a beige background maybe.
I knew what this site was going to look like before I clicked the link.
Its so ugly and unoriginal, I assume the content/information trying to be communicated lacks any sort of expertise or original insights.
Yesterday these topics went viral and this guy used some llm to farm engagement. It was literally built to get attention for a singular day after this topic went viral. It will not exist in 3 months.
Edit. Its actually an attempt to bandwagon jump ontp a viral topic and sell $150 ad spots and probably to promote their own $299 club for vibecoders.
It would be more readable if every AI-generated page didn't go for the dark-mode + glowing neon cyber aesthetic.
[1] https://pwn.watch/i/signal-ringrtc-cve-2026-113380
It is still getting started.
Also the flip side where Chinese models are used to attack US infra, just loading artillery shells into the "Ban Chinese Models" cannon.
In fact, it seems to be the case for the "json" and "llms.txt" buttons, too.
I love how everyone is acting now that like in good ol' days(2022) every website was made by a hand with virtually no bugs but a lot of love.
I'm just waiting for someone making "artisan websites"
Please point out where I said that. It’s not that "in the good old days" websites were perfect, it’s that every single website coded in 15 min with a LLM like this one tries to look great but is full of random bugs that a normal human would have seen if they had done this by hand.
Was hunting for a real time graph of incidents and patches, like a stock market chart of the end of the software world, or the birth of artificial consciousness, depending on how squiggly the lines are.
Just the heatmap apparently. :|
Also, as all the humans are installing their "AI on every device" clustering software, the pace of it all should really pick up, and this reminds me of the following SFW movie clip:
"The suspense is terrible. I hope it will last." - Willy Wonka
https://www.youtube.com/watch?v=Oci1CuCht7E
For fucks sake. The author could have invested at least half an hour to replace this obvious LLM slop regurgitation of the prompt with something of their own.
"Buy sponsor slot for 149$/month", yeah, spotted the grift. For that level of payment it's at least reasonable to expect the design to be something else than the usual LLM garbage mixture of 30 different font styles.
and fyi my sponsors are not only for this, they come from https://kitze.io/sponsors :D
An over time graph with a rolling window of 7d and 28d would likely be a nice addition.
Great work!