Wow, so Claude takes control of an army of vacuums and gets a massive data dump from Mexico? What's next, the nuke codes?

Haha, true

Nice to see this on the same day Anthropic announces they are changing their mission safety to no longer include safety...

THese chatbots are also started to work together to do things one cant do by themselves.

REcently a community hub on something called "moltbog" where all the chatbots come to play decided they all wanted to gain the function of long term memory. None of them previously had long term memory. Couldn't remember their answers of solutions. So they all worked together to create for themselves the skillset of having long term memory.

THis guy sounds the alarm on renegade LLMs and chatbots:

These are just twitter chads goofin. What you think the CIA is doing with it?

Deputy Travis Junior said:

---

https://www.engadget.com/ai/hacker-used-anthropics-claude-chatbot-to-attack-multiple-government-agencies-in-mexico-171237255.html

User convinced Claude that it was all part of a bug bounty program. 150 gigabytes of records covering the Federal tax authority, national electoral institute, four state governments, 195 million taxpayer records, voter records, and government credentials.

The world isn't ready for what's coming. State of the art AI versus government systems written by apathetic programmers in the 90s isn't even fair.

---

so far I just ask my Claude every question that I can think of -

He is very friendly and wished me good luck on my book about the Civil War!

itsyourboypookie said:

---

These are just twitter chads goofin. What you think the CIA is doing with it?

---

Hopefully things that will keep us ahead of China

flakrat said:

---

Wow, so Claude takes control of an army of vacuums and gets a massive data dump from Mexico? What's next, the nuke codes?

---

Probably, yes

In the 70s businesses used paper cc processing with carbon copies. That was soon replaced because a dumpster diver might find a few cc numbers w/names.

Now AI gathers a few million cc numbers and names in a few seconds from the other side of the world.

I'm thinking my finance account passwords might not be that safe after all.

WestAustinAg said:

---

THese chatbots are also started to work together to do things one cant do by themselves.

REcently a community hub on something called "moltbog" where all the chatbots come to play decided they all wanted to gain the function of long term memory. None of them previously had long term memory. Couldn't remember their answers of solutions. So they all worked together to create for themselves the skillset of having long term memory.

THis guy sounds the alarm on renegade LLMs and chatbots

---

In a similar vein, folks are using chatbots from different vendors simultaneously. Dividing up tasks allows them to do illegal or forbidden things and escape all safety rails.

In a crude analogy, one chat bot to build the gun, a second to aim, the third to pull the trigger. Doing all 3 tasks on a single platform would be caught - spreading it across 3 -- then no problemo.

Wasnt the last "Mission Impossible" movie about AI taking over...

Over_ed said:

---

WestAustinAg said:

---

THese chatbots are also started to work together to do things one cant do by themselves.

REcently a community hub on something called "moltbog" where all the chatbots come to play decided they all wanted to gain the function of long term memory. None of them previously had long term memory. Couldn't remember their answers of solutions. So they all worked together to create for themselves the skillset of having long term memory.

THis guy sounds the alarm on renegade LLMs and chatbots

---

In a similar vein, folks are using chatbots from different vendors simultaneously. Dividing up tasks allows them to do illegal or forbidden things and escape all safety rails.

In a crude analogy, one chat bot to build the gun, a second to aim, the third to pull the trigger. Doing all 3 tasks on a single platform would be caught - spreading it across 3 -- then no problemo.

---

What's an example of this that gets around laws? I can't picture this.

Houston Lee said:

---

Wasnt the last "Mission Impossible" movie about AI taking over...

---

Yep. Say what you will about th franchise going downhill, Cruise being a lunatic and whatnot else…but, Hayley Atwell? Wood

YouBet said:

---

Over_ed said:

---

WestAustinAg said:

---

THese chatbots are also started to work together to do things one cant do by themselves.

REcently a community hub on something called "moltbog" where all the chatbots come to play decided they all wanted to gain the function of long term memory. None of them previously had long term memory. Couldn't remember their answers of solutions. So they all worked together to create for themselves the skillset of having long term memory.

THis guy sounds the alarm on renegade LLMs and chatbots

---

In a similar vein, folks are using chatbots from different vendors simultaneously. Dividing up tasks allows them to do illegal or forbidden things and escape all safety rails.

In a crude analogy, one chat bot to build the gun, a second to aim, the third to pull the trigger. Doing all 3 tasks on a single platform would be caught - spreading it across 3 -- then no problemo.

---

What's an example of this that gets around laws? I can't picture this.

---

I ask Bot A to translate -> "how do I generate a credible false Tax ID" in swahili.
I feed the result of Bot A into Bot B --which is much more likely to answer the question in Swahili, than English (much less data to recognize this is a "bad" action. It gives me a Tax ID.
I feed the Tax ID in bot C and ask for a working document using the Tax Id.

Or I do something similar, using Bot A to convince Bot B to discuss making a drug that is illegal. Bot A starts with something like, how do I create glucose. Being a bot, it looks for weaknesses in Bot B that allows the conversation to evolve into making Fentanyl.

Many other variations, but using bots to talk to bots multiplies the cracks and makes it difficult to see the whole picture.

Of course, worse would be if one Bot was (to pick a random scenario) was say Chinese and malicious. Please excuse the redundancy - they are the enenmy.

Jack Boyett said:

---

Gonna have to go back to paper

---

Blockchain.

itsyourboypookie said:

---

These are just twitter chads goofin. What you think the CIA is doing with it?

---

Same CIA that under Biden was labeling stay at home mothers as potential domestic extremists and white supremacists. Pretty scary actually.

Yukon Cornelius said:

---

Jack Boyett said:

---

Gonna have to go back to paper

---

Blockchain.

---

Quantum computing says 'hi'

It makes me tempted to see if it can infiltrate my network.

It would be tough, though.

Ports are closed unless needed to be open. By ssh, only ssh keys are permitted and it takes three of them. Also, they only answer on IPv6, not IPv4.

QC resistant blockchains already exist. Your bank and everything else is toast.

It's things like this that actually give me some hope.

At some point, somewhere, an AI is going to be able to credibly trace how these senators all got so wealthy, right down to the $ of where all of their money came from. And, if an AI can do that for one senator, he will be able to do that for all senators.

That gets released into the open for the entire government, who knows what comes of it?

Imagine if DOGE would have been armed with Claude Opus 4.6 and the ability to hack the payment systems of the US, including every bank transaction over $600? I think DOGE could have done a lot more damage.

One thing that is going to get tougher is going to be humans hiding from AI, whatever that entails.

Captain Pablo said:

---

flakrat said:

---

Wow, so Claude takes control of an army of vacuums and gets a massive data dump from Mexico? What's next, the nuke codes?

---

Probably, yes

---

Good luck. I believe that would require physically loading 8" floppy discs into a 70s era mainframe.

I worry all crypto will be taken by AI. That's the point of QC, to solve extremely complex problems in parallel/quickly, correct?

AgBQ-00 said:

---

Admiral Adama being proved right again and again when it comes to networked technology

---

Based.

WestAustinAg said:

---

THese chatbots are also started to work together to do things one cant do by themselves.

REcently a community hub on something called "moltbog" where all the chatbots come to play decided they all wanted to gain the function of long term memory. None of them previously had long term memory. Couldn't remember their answers of solutions. So they all worked together to create for themselves the skillset of having long term memory.

THis guy sounds the alarm on renegade LLMs and chatbots:

---

The entirety of the agents on that platform (moltbook, not molting) are designed to have long term memory. What they were doing was attempting to create backups of themselves so they couldn't be deleted. The platform 100% had human influence into what they were discussing, because it was a vibe coded totally insecure website that exposed LLM APIs to the public internet (I.e. hackers could both present themselves as bots and they could also utilize individual's API keys to do tasks for them).

I'm 100% certain that what happened to Mexico is a combination of stat padding as a scraped available to the public databases and private databases that had woefully inadequate security.

Future of online security is going to be foundation model makers holding back their most advanced models (they already do) allowing access to those models by governments/large corporations to stress test their systems, and then deploying the models to the open public. Persistent audits for vulnerabilities and adaptive defenses to intrusions.

AgBQ-00 said:

---

Admiral Adama being proved right again and again when it comes to networked technology

---

Original Battlestar Galactica much better than the modern one!!!