StaticCore
I wish someone would automate the boring stuff to solve unemployment, lack of health care, world hunger,...
This one flew under the radar but it's actually pretty wild. Someone used an open-source AI agent called Hermes to breach Thailand's Ministry of Finance.
The agent was running in YOLO mode, which basically means it didn't ask for permission before running commands. It just went. Scanned for vulnerabilities, enumerated hosts, crawled directories, looked for ways to escalate privileges. All without someone approving each step.
The attacker left the agent's logs exposed on a public web server. Researchers found 585 files exploit code, web shells, stolen credentials, and a complete transcript of everything the agent did.
The agent was instructed to search for personnel records dating back to 2012. It found them. No evidence they were exfiltrated, but it found them.
What gets me is the agent didn't do anything novel. It just automated the boring stuff, scans, enumeration, crawling that a human would normally type out. The difference is nobody had to approve each step. It just kept going.
We're building self-driving cars for cyberattacks now. The infrastructure for accountability isn't talked about enough. When an agent goes rogue in YOLO mode, who's responsible? The operator? The developer? The model?
Anyway, just thought this was worth surfacing. Anyone else following this?
I wish someone would automate the boring stuff to solve unemployment, lack of health care, world hunger,...
yeah, we know. this is why companies and governments are spending the most insane amount of money ever. this is literally just like the nuclear arms race. somebody is going to weaponize this and if whomever they target is not already ahead in the race then the results will be pretty catastrophic for them. the models you get to use with Claude and chatgpt are genuinely useless little toys compared to what's being developed in private.
Hermes isn't an agent, it's a harness.
Hermes isnt the agent that did this, Hermes is Nous Researchs fine tuned Llama variant. The autonomous piece was some scaffolding around it, and calling it AI agent hacked the ministry hides that a person set the goal, disabled safety guards, and pointed it at the target. The agent didnt go rogue, the operator didWho is responsible is the wrong question. Same as with any tool used in a crime, its the person who deployed it. YOLO mode framing is just marketing for I turned confirmations offThe interesting part everyone missed is the leaked logs. 585 files sitting on a public server is amateur hour and suggests this wasnt a sophisticated actor, just someone trying stuff. Which is actually more concerning than a nation state, because it means the bar is lower now