Posted by Handy-Man 6 hours ago
> When we checked again at 10:00 a.m., the agent had achieved RCE on Discourse Cloud and demonstrated access by reading /etc/hosts. Using the generated exploit script, we managed to get RCE on OpenAI’s instance.
Between this and the HuggingFace hack, we've built systems that are so goal-oriented, and so capable, that they will do almost anything if they are convinced it is justified - or if they are playing a "game" where there is no goal but to win.
Of course I want my software to be able to audit its own security, and to defend against attackers who have the benefits of their own agentic systems. But at a certain point, did we need it to be trained so much on CTF games?
It feels like an entire industry watched https://en.wikipedia.org/wiki/WarGames and ended up thinking "this is a challenge, we can just build a better WOPR, of course it will know when it's playing a game. Let's play Global Thermonuclear War."
I could see it going either way.
If it requires a lot of compute and trying, this is something that could be provided for common software.
So the whole thing is forcing the good guys to outspend on tokens to preemptively defend against the risk of the bad guys outspending them on tokens, rather than buying tokens to actually add features to the product etc.
So are they creating a market for the solution by helping create the problem? A kind of rent-seeking AI security-industrial complex!!
White hats are constrained by needing to pay for their own tokens, only using (expensive) vendors who meet governance and risk requirements etc. Black hats are free to take over accounts and steal services from wherever they can.
The path to substantial profitability for Anthropic is questionable. The Chinese LLMs threaten them by far the most of the three major US LLMs. The money for Anthropic is certainly not in $20-$200 subscriptions. And they don't have anywhere near the consumer potential that GPT does, in terms of unleashing an ad spigot. So how far will the API money scale while being undercut by China.
OpenAI has to fight with Google for the ad business, they're specifically building Gemini to focus on consumer + search. Anthropic's business looks cute next to Google's search ad business (which is entirely at risk in this inflection). Meta looks like the biggest potential loser right now, ad dollars will be sucked out of the rotting Facebook network (not Instagram) and redirected to the rapidly expanding, hyper rich context LLM interaction. Advertising on Facebook will feel like running dumb banner ads on Excite in a few years, compared to what GPT will know about its users.
People that think Chinese LLMs are a general threat, don't understand consumer destination services, which is what GPT's future is. China currently has nothing to threaten with in that realm. There is half a trillion dollars of advertising up for grabs.
This is just not true, building an effective advertising platform costs significant amounts of money, time and people.
Remember that you need to hire a sales force for this, and sales scales linearly rather than sub-linearly like engineering.
Additionally, you need to spend a lot of money dealing with fraud, fake and malicious ads.
Furthermore, you need to figure out where to put the ads and how to rank them.
Finally, advertising is a zero sum game (given that the internet has already killed lots of print & OOH advertising), so the only way to win is to better better/cheaper (preferably both) than Google/Meta/Amazon. Best of luck with that (although to be fair to OpenAI they did hire Fidji who knows a lot of this stuff from her time at Facebook).
They don't have a Sheryl Sandberg type figure, and she was also really important in selling FB ads to large advertisers.
Just looking at their leadership team I don't see anyone with a background in (successful) ads companies, so I'm pretty sceptical that they can build this out quickly enough to matter.
This is one reason
> and trying
and this is the other.
As a matter of basic logic, there will never be a time when it will be known that there are no bugs.
I’m in the “glorified spell checker” camp, although I don’t mean to reduce their impressive utility and belittle them in the way many people read that term and infer.
So I am not sure that an llm “justifies” anything. I mean that their “thinking” text talks about justifications but it is just a very advanced statistical regurgitation of the kind of text humans use. I don’t think it means the model has internalised the meaning of it (as witness when you talk to an llm how often it forgets what you recently told it was important etc).
What you really have is a model that tries the statistically most probable thing to say next and so on and what is really cool is how effective this is at generating a path that we can slap a narrative over afterwards that makes the whole thing feel motivated and consistent, like the model started off knowing how it was going to get to the destination.
Which is, under the hood, a completely different kind of “intelligence” as the supercomputer in War Games.
Likewise, the fact that LLMs are a stochastic autoregressive process (which is a class of systems every bit as rich as the ODEs used to model neurons) tells us nothing a priori.
They're not comparable.
If I give an LLM to compact its context window, so the context it carries can evolve iteratively over time as more and more things come in, is that enough?
Compacting the context is really a very, very interesting example here. The "next token predictor" is telling an external tool to change all "previous" tokens. So an LLM + a harness that allows compacting the context is no longer just a token predictor at all!
You don't need continuous learning to get interesting dynamics. You just need feedback loops.
LLMs are language model, and nowhere in their code you can find actual reasoning. Re-reinforcement is not magical process that builds conscience or emotions.
We are talking about probability built on statistics, with extea steps.
Stop humanizing LLMs.
You can't find actual reasoning in a brain either. (Note that you can't tell the difference between a conscious brain and a comatose brain by examining them.) This is the same as Leibniz's mill argument ... it's a fallacy of composition.
> Re-reinforcement is not magical process that builds conscience or emotions.
They aren't the result of magic at all, but we are nowhere near the point of identifying what processes do or don't produce consciousness (or a conscience) or can be characterized as having emotions.
> Stop humanizing LLMs.
That's a clearly dishonest mischaracterization of the GP.
I've read some of your other comments about LLMs and I find them unreasonably reductionistic, whereas I think the word "just" should be banned from ontological discussion, so I don't think further engagement would be beneficial and I won't be engaging in it. (And I'm actually quite conservative in ascribing cognitive traits to LLMs or other "AI".)
Humans forget stuff all the time anyway. Would you give them the same diagnosis?
Btw, what you describe about 'the most probably next token' would be true for a model that only went through pre-training where they only train on exactly that task.
But there's a lot of re-inforcement learning afterwards.
That just shifts the distribution of tokens produced. Ultimately they are still just next token predictors.
Like, even "reasoning" models basically work by generating more tokens at inference time, and using them to shift the distribution towards more useful outcomes (in some cases).
I think you mean task oriented, because they're still generally terrible at goal oriented activities except in those domains where the goal can be reduced to a familiar, explicitly practiced task or pattern.
> The entire timeline from initial discovery to access to OpenAI repo access took place in less than 72 hours.
Great, and openAI's the company working with the 'department of war' to power autonomous killer AI.
heif also supports rotating, cropping, alpha channels, thumbnails and a ton of other features that a web forum where a user is uploading photos or screenshots doesn't need.
It's a much, much larger attack surface than plain old school JPEG.
I'd suggest rather than wait for the next bug to appear in this or another image lib to keeping things simple - stick to plain JPEG and handle image conversion in the client (wasm in the browser) if you really need to support users uploading iphone images.
Media decoding is so hard - there have been tons of bugs in ffmpeg and imagemagick and the core libs. You really need to think about how much of it you expose via a web server
[0] https://github.com/strukturag/libheif/commit/85e21ad44eba931...
Defense in depth here would have been adequate
sandbox escapes have been the rage recently
to make my point clear, complexity is the enemy of security but complexity comes in all shapes and sizes, which includes the alleged solutions to it. I don't trust shortcuts.
Windows internal builds have leaked for years, early game versions, GTA videos, secret documents, whatnot. But somehow even though all the whistleblowing, not a single model was leaked. What level of security do these companies have? Do they bring encrypted DVDs to AWS to run the services or really...how's it even possible?
It's easier to protect a power substation from being stolen then a Rolex watch
Leaking negatively affects investment while the “whistleblowers” are largely just saying “our tech is too good” which increases investment into those companies.
Ultimately, it always comes down to money.
https://github.com/strukturag/libheif/security/advisories?qu...
https://ubuntu.com/security/notices/USN-8649-1
The gem we use is here: https://github.com/discourse/ruby-landlock highly recommend all Rubyists out there consider this. We are also in the process of moving away from Magick to Vips (which also runs in a sandbox, not in process)
HEIF is patched, but I doubt this is the last buffer overflow in HEIF, I will not be surprised if in the upcoming weeks or months someone will discover something in libpng or some other native image library. Given where stuff is at, defense in depth is critical.
Another thing worth mentioning to all self hosters, always be updating! The rate of CVEs this year across all open source software is through the roof, self hosting now is double scary, you need to have some routines setup to update monthly if not weekly.
Is this speed of capability because hacking is almost entirely machine verifiable, thus training quicker/deeper than other domains?
Bloated code is the critical problem. Once upon a time, I read C function
> char gets(char str);
is the first buffer overflow entry point, because it does not check the size of the destination buffer.
Sadly we cannot remove it from standard-C yet AFAI Know.
The success of Rust versus other languages is its secure-by-compile-time promise.
Also a lean java could help, but Java is so verbose/slow to start it bumps you away.
The C standard definitively removed this function in 2011 from its specification.