To some degree that's just how the English language works. "to hide" can be a passive effect (the hedge hides a garden) or an action with intent (the robber hides the weapon). Usually we distinguish between them based on whether the subject of the sentence has agency. But that breaks down with LLMs. Same as in sentences about the brain
There's no obscurity anymore. Best you can do is keep as much as possible server side. Curious if anyone has experimented with LLM's RE'ing obfuscated code. I suspect they handle it just fine.
RE was arguably the first* task category to become completely solvable by LLMs on cybersecurity CTFs. This is exactly about obfuscated code, LLMs seem to have no problems with it now. The only probabilistic defense is to include tons of red herring so agent goes the wrong path and ends up in some tarpit like bruteforcing unrelated un-bruteforceable hashes.
* Cryptography category also went down fast but I think LLMs solved RE a bit faster. RE and crypto have in common that the task is completely local, but in classic RE tasks it is also usually ~obvious what you are supposed to do, you just need to do it - which agents do really well.
My layman understanding (as someone who isn't a lawyer but has been tasked with some reimplementations), is that it at least is governed by different sections of law. For me it's the European Software Directive where Article 5 is pretty generous and governs basically "observing" software for reverse engineering. Article 6 is less generous and is about decompilation. There you basically can only to it to ensure interoperability.
So depending on what the goal is (making an interoperable piece of software, or making a clone?), you might want to stay under Article 5, and steer clear of Article 6.
> Curious if anyone has experimented with LLM's RE'ing obfuscated code.
Sort of, albeit I haven't checked with SotA obfuscation, I've seen GPT 6.1 Sol find a URL hidden under a xorpad + 2 layers of encryption, and I've seen Astra RE an AES whitebox.
More importantly, remember these game decomps (e.g. Pokémon)? Well, with minimal guidance and IDA MCP these new models are able to do this _and_ properly name all function/params/locals in a single day.
today i read that for new game studios, the possibility of going stream only (like amazon luna) instead of downloadable is a probable future if they want to prevent piracy.
Some of it is. He did the classic “book an entire vacation for me“ except he only asked it to plan.
No one is ever going to just tell a phone to book a vacation for them. It won’t know what you want well enough. Planning is much more reasonable because then you can just adjust whatever you don’t like.
But as a demonstration, it certainly shows how far Siri has come. It probably would’ve just offered to do a web search before.
The things I’ve found it really useful for so far is the “find this piece of information that must exist on my phone somewhere, but I don’t have the slightest clue where“ kind of query.
Would love to learn more about some techniques that "everybody" uses to do this well. So far, everything I've seen that meaningfully advances the frontier has been high-touch (involving human experts in one way or another).
It's fairly easy to describe a task that is slightly harder than an existing one.
For example if frontier models are able to one-shot a database query across 20 columns and 10 tables add one additional relationship then test. Keep doing this until the pass-rate drops below acceptable and now you have your new frontier eval.
My thought experiment was along the lines of "Let's say I'm Anthropic and I want to significantly improve my frontier model's performance on, say, theoretical physics research. How do I build a fully autonomous process capable of constructing an eval that's somewhat outside the current capability in some useful direction (decided by the autonomous process itself)?"
This hits. Ive had an annoying issue with Codex VS Code extension for weeks if not months when queuing up a follow up. Often it just disappears, I think it's just a UI glitch is actually queued but I can't click steer if the UI doesn't update and I find myself automatically going CMD+A, CMD+C before sending a follow up message.
reply