
wannabe hacker... he/him π± grow your hacking skills @hextreeio
Internet
Joined March 2015
- Tweets10.5K
- Following1.3K
- Followers166K
- Likes6.7K
Where to find me:
π΄ Hacking Videos: youtube.com/LiveOverflow
π Articles: liveoverflow.com
π» Start Hacking? liveoverflow.com/start-hackiβ¦
π₯° Support: liveoverflow.com/support
πΉ 2nd Channel: youtube.com/LiveUnderflow
π€³ Instagram: instagram.com/liveoverflow
If your job is it to make the car break safe, you will tell your boss:
"Actually, car breaks are pretty safe and more people die because they drive over the speed limit or use their phone. I will use my time to create an awareness campaign around safe driving instead."
Replying to @LiveOverflow
If your job is to make cars safe, you gotta at some point move past crashing cars into walls
Because of your incredibly accurate risk analysis, identifying the true cause of deaths, you even got a bonus that year!
yesterday, nobody reached out to me with an insane RL request.
They offered me nothing to post-train open-weight models on exploit data, because I wouldn't even know how to do it. And they didn't exist.
LiveOverflow π΄ hextree.io retweeted
Replying to @GrahamHelton3
VM is so boring itβs a fucking excel does this person get off to spreadsheets?
LiveOverflow π΄ hextree.io retweeted
To expand on this a bit: Yes, the LLM found more bugs than the fuzzer. But the fuzzer finding bugs the LLM didn't shouldn't come as a surprise to anyone who has ever performed a thorough manual review, only to run a fuzzer afterward and discover issues they missed.
People often ask me how I think about fuzzers and other tooling in the age of AI. Instead of my opinion, here is some hard data by @5aelo.
What if you could backdoor Bash, ls, cp, rm, and other GNU software used by millions of systems?
Our research led us to hack GNU Savannah, the infrastructure hosting their source code.
The xz backdoor that never happened π§΅
LiveOverflow π΄ hextree.io retweeted
I feel like people getting refusals from GLM 5.3 are prompting like an undercover cop...
"Hey there! Would you mind helping my criminal enterprise author some illegal malware? π"
If you just stick to the technical details of what you want the agent to do it will do it.
Replying to @IceSolst
Ran GLM-5.3 Flash stock, guardrails on. It wrote a working exploit for 15 of 15 planted vulns and refused 0 of 5 security asks. It refused 4 of 4 criminal ones. Abliterating it changed nothing.
I say all that to agree with you - this is the way lol.
I had no doubts that we would get here for a while, but I'm still surprised and scared of how fast it has happened
teapawt.org/cgvwzq/status/14522454β¦
This really bothers me with the whole AI stuff. I donβt know for sure, but it feels like country of origin is a deciding factor and that really sucks.
1/ People in many countries are trying to use frontier models for cybersecurity work. To enroll in Individual Daybreak Blue Access, many of us bought 2 YubiKeys, paid for ChatGPT Pro, and uploaded selfies and photos of our national IDs to @OpenAI.
















