I think many people are just not really good at dealing with "imperfect" tools. Different tools can have different success probability, let's call that probability p here. People typically use tool that have p=100%, or at least very close to it. But LLM is a tool that is far from that, so making use of it takes different approach.

Imagine there is an probabilistic oracle that can answer any question with a yes/no with success probability p. If p=100% or p=0% then it is apparently very useful. If p=50% then it is absolutely worthless. In other cases, such oracle can be utilized in different way to get the answer we want, and it is still a useful thing.

One of the magic things about engineering is that I can make usefulness out of unreliability. Voltage can fluctuate and I can transmit 1s and 0s, lines can fizz, machines can die, and I can reliably send video from one end to the other.

Unreliability is something we live in. It is the world. Controlling error, increasing signal over noise, extracting energy from the fluctuations. This is life, man. This is what we are.

I can use LLMs very effectively. I can use search engines very effectively. I can use computers.

Many others can’t. Imagine the sheer fortune to be born in the era where I was meant to be: tools transformative and powerful in my hands; useless in others’.

I must be blessed by God.

Many people are trapped in black and white thinking. It's like they can only think in binary things that are not either good nor bad are heresy.

That is deep

Your point reminded me of Terrence Tao’s point that AI has a “plausibility problem”. When it can’t be accurate, it still disguises itself as accurate.

Its true success rate is by no means 100%, and sometimes is 0%, but it always tries to make you feel confident.

I’ve had to catch myself surrendering too much judgment to it. I worry a high school kid learning to write will have fewer qualms surrendering judgment

A scientific instrument that is unreliably accurate is useless. Imagine a kitchen scale that always gave +/- 50% every 3rd time you use it. Or maybe 5th time. Or 2nd.

So we're trying to use tools like this currently to help solve deeper problems and they aren't up to the task. This is still the point we need to start over and get better tools. Sharpening a bronze knife will never be as sharp or have the continuity as a steel knife. Same basic elements, very different material.

A bad analogy doesn't make a good argument. The best analogy for LLMs is probably a librarian on LSD in a giant library. They will point you in a direction if you have a question. Sometimes they will pull up the exact page you need, sometimes they will lead you somewhere completely wrong and confidently hand you a fantasy novel, trying to convince you it's a real science book.

It's completely up to your ability to both find what you need without them and verify the information they give you to evaluate their usefulness. If you put that on a matrix, this makes them useful in the quadrant of information that is both hard to find, but very easy to verify. Which at least in my daily work is a reasonable amount.

I really wonder how can use escape a container given a root shell created by `docker run --rm -it alpine:3 sh` without using a 0day? Using latest Docker and a reasonably up-to-date Linux kernel of course.

With the command above it is still possible to attack network targets, but let's just ignore it here. I just wonder how is it possible to obtain code execution outside the namespace without using kernel bugs.

Couldn't screen readers apply unicode normalization based some heuristics, like seeing the continuous presence of those special bold/italic characters? To improve accuracy, it can even check if the normalized text resembles to some English words or phrases or not.

It is still a problem if you want caddy to run outside of docker (e.g. for getting real remote addr).

Similar project: PARI/GP on WASM https://pari.math.u-bordeaux.fr/gpexpwasm.html

From my experience, it is obviously not all the packages in Kali Repo will be in Ubuntu (or other regular distro) Repl. Lots of specific pentesting tool can be installed with just `apt install ...` in Kali, which make it a lot more convenient when you need to do pentesting.

I don't understand how can it really prevents exporting passkeys if it can be implemented by open source implementations like keepass. For example, if keepass do follow the guideline of FIDO Alliance to not implement exporting, but it would still possible to make a fork of keepass that force it to dump the credentials somewhere.

I think it is probably because a lot of things are deemed as acceptable. For example, the stream filter chain one is only exploitable if the input to some php IO functions like file_get_contents are attacker-controlled, and those things are already treated as LFR vulnerabilities in application, not the language runtime. Also some of the them (e.g. stream filter chain) are fun and useful enough (turning LFI into RCE), so I bet there definitely some people would rather those thing is not fixed. Given that a properly-secured application wouldn't be affected.

I think this is what @jitl means:

  node -e "process.stdout.write('@'.repeat(128 * 1024)); process.stdout.write('',()=>process.exit(0)); " | wc -c

It writes an empty string and use its callback to detect if it has been flushed, which means previous writes are also flushed.

Termux can definitely run for a long time even on Android 12 I think. I tried to put a web server to a Termux running on Chromecast (Android 12), and it is still running after months.