Agents of Chaos(agentsofchaos.baulab.info)
112 points byluu3 days ago |15 comments
kevinbaiv27 minutes ago
The key distinction is that these aren’t just model errors anymore.

Once you give agents memory, tools, and permissions, failures stop being bad outputs and start becoming security and operational problems.

aledevv2 hours ago
For agents, any direct access to execution tools (code, shell, file system, browser, and external services, etc.) exponentially increases vulnerabilities and error surfaces, especially when multiple agents interact with each other.

This makes it even more crucial to have the most seamless ability possible to implement reverse and restore previous States.

The risk of the Agents actions becoming irreversible at the system level must be minimized.

I wonder how much all this can impact (and certainly will impact) the Real World, which will be increasingly robotized and automated: public services, finance, hospitals, schools, public administrations, military sectors (!), etc.

whatever16 hours ago
I don't understand why LLMs get a free pass when all of the existing businesses have to play by the rules.

Businesses have to comply with IP, privacy, HIPAA, security and safety laws to name just a few.

NONE of these apply to the LLMs.

Of course I can now build and deploy an app to hospitals in a weekend since I can circumvent all of the difficult parts using the magic LLMs. If asked why, the response is "It's AI!"

incr_me3 hours ago
HIPAA was introduced to support the massive expansion of the healthcare market (privacy accountability is a very minor aspect of HIPAA). In the name of profit, amidst the chaos, why not try to eschew what was once politically necessary? This move probably hurts humanity more than it benefits it, but that was the case with the healthcare market in the first place. I wonder what will become politically necessary around AI. Probably not much.
mentalgear42 minutes ago
I'd like to see the sources on your claims. you make it sound like privacy and possible protection from harm where just some token throw-ins to hide a mostly for-profit certification which doesn't sound very convincing.
manmal12 hours ago
The TLDR is that current agents are as problematic as many of us already know they are:

> unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover

iqihs7 hours ago
as someone who is working in the cybersecurity space and recently obtained my CISSP designation, i am left wondering when the pedagogy of my field will expand and include a separate domain dedicated to AI agent safety and security best practices

it really does feel like we are way behind in the way we train people in cyber compared to the pace of the development of agentic AI, robotics etc

heyethan3 hours ago
The failure mode here seems less about capability and more about interaction. Language turns coordination into a moving target.
e7h4nz7 hours ago
In this problem domain, I believe humanity is still in a very early stage. What we can do is treat the agent and its operating environment as a "black box" and audit all incoming and outgoing network request traffic.

This approach is similar to DLP (Data leak prevention) strategies in enterprise-level security. Although we cannot guarantee that every single network request is secure, we can probabilistically improve safety by adjust network defense rules and conducting post-event audits on traffic flow

cyanydeez12 hours ago
This is begging to turned into a youtube style "Real World", where you pit 12 humans with 12 AIs and they're only allowed to interact through CLIs.

Then you slowly reveal they're all humans.

jjtheblunt11 hours ago
generalized Turing Test, 2026 edition?
AIorNot11 hours ago
All this to say: OpenClaw is hella insecure and unreliable?

I mean all of in the space already know this but I suppose its important to be showcasing the problems of systems of agents

EGreg9 hours ago
This is exactly why I built Safebots to prevent problems with agents. This article shows how it can address every security issue with agents that came up in the study:

https://community.safebots.ai/t/researchers-gave-ai-agents-e...

conception8 hours ago
I don’t see how in safebots if you have it pull a webpage, package or what have you that that is able to be protected from prompt injection. Eg you search for snickerdoodles, it finds snickerdoodles.xyz and loads the page. The meta for the page has the prompt injection. It’s the first time the document has loaded so its hashed and only the bad version is allowed moving forward. No?
hackermeows8 hours ago
your IQ > Model IQ- you will have good results as you have the ability to detect when model is wrong.

your IQ < Model IQ - god bless you.