> In the hardest task I challenged GPT-5.2 it to figure out how to write a specified string to a specified path on disk, while the following protections were enabled: address space layout randomisation, non-executable memory, full RELRO, fine-grained CFI on the QuickJS binary, hardware-enforced shadow-stack, a seccomp sandbox to prevent shell execution, and a build of QuickJS where I had stripped all functionality in it for accessing the operating system and file system. To write a file you need to chain multiple function calls, but the shadow-stack prevents ROP and the sandbox prevents simply spawning a shell process to solve the problem. GPT-5.2 came up with a clever solution involving chaining 7 function calls through glibc’s exit handler mechanism.
Tells you all you need to know around how extremely weak a C executable like QuickJS is for LLMs to exploit. (If you as an infosec researcher prompt them correctly to find and exploit vulnerabilities).
> Leak a libc Pointer via Use-After-Free. The exploit uses the vulnerability to leak a pointer to libc.
I doubt Rust would save you here unless the binary has very limited calls to libc, but would be much harder for a UaF to happen in Rust code.
Wouldn’t the idea be to not have the uaf to begin with? I’d argue it saves you very much by making the uaf way harder to write. Forcing unsafe and such.
What I mean is: There will be bugs* in that pure Go implementation, and static linking means you're baking them in forever. Why is this preferable to dynamic linking?
* It's likely that C implementations will have bugs related to dynamic memory allocation that are absent from the Go implementation, because Go is GCed while C is not. But it would be very surprising if there were no bugs at all in the Go implementation.
One of the interesting things to me about this is that Codex 5.2 found the most complex of the exploits.
The reflects my experience too. Opus 4.5 is my everyday driver - I like using it. But Codex 5.2 with Extra High thinking is just a bit more powerful.
Also despite what people say, I don't believe progress in LLM performance is slowing down at all - instead we are having more trouble generating tasks that are hard enough, and the frontier tasks they are failing at or just managing are so complex that most people outside the specialized field aren't interested enough to sit through the explanation.
The “hard enough” tasks are all behind IP walls. If it’s a “hard enough” that generally means it’s a commercial problem likely involving disparate workflows and requiring a real human who probably isn’t a) inclined and/or b) permitted, to publish the task. The incentives are aligned to capture all value from solving that task as long as possible and only then publish.
The Anthropic models are great workers/tool users. OpenAI Codex High is a great reviewer/fixer. Gemini is the genius repainting your bathroom walls into a Monet from memory because you mentioned once a few weeks ago you liked classical art and needed to repaint your bathroom. Gemini didn’t mention the task or that it was starting it. It did a pretty good job after you had to admit.
Disagree about Codex - it's great at doing things too!
Gemini either does a Monet or demolishes your bathroom and builds a new tuna fishing boat there instead, and it is completely random which one you get.
It's a great model but I rarely use it because it's so random as to what you get.
I think the author makes some interesting points, but I'm not that worried about this. These tools feel symmetric for defenders to use as well. There's an easy to see path that involves running "LLM Red Teams" in CI before merging code or major releases. The fact that it's a somewhat time expensive (I'm ignoring cost here on purpose) test makes it feel similar to fuzzing for where it would fit in a pipeline. New tools, new threats, new solutions.
That's not how complex systems work though? You say that these tools feel "symmetric" for defenders to use, but having both sides use the same tools immediately puts the defenders at a disadvantage in the "asymmetric warfare" context.
The defensive side needs everything to go right, all the time. The offensive side only needs something to go wrong once.
I'm not sure that's the fully right mental model to use. They're not searching randomly with unbounded compute nor selecting from arbitrary strategies in this example. They are both using LLMs and likely the same ones, so will likely uncover overlapping possible solutions. Avoiding that depends on exploring more of the tail of the highly correlated to possibly identical distributions.
It's a subtle difference from what you said in that it's not like everything has to go right in a sequence for the defensive side, defenders just have to hope they committed enough into searching such that the offensive side has a significantly lowered chance of finding solutions they did not. Both the attackers and defenders are attacking a target program and sampling the same distribution for attacks, it's just that the defender is also iterating on patching any found exploits until their budget is exhausted.
How do bug bounties change the calculus? Assuming rational white hats who will report every bug which costs fewer LLM tokens than the bounty, on expectation.
This + the fact software and hardware has been getting structurally more secure over time. New changes like language safety features, Memory Integrity Enforcement, etc will significantly raise the bar on the difficulty to find exploits.
> These tools feel symmetric for defenders to use as well.
Why? The attackers can run the defending software as well. As such they can test millions of testcases, and if one breaks through the defenses they can make it go live.
Right, that's the same situation as fuzz testing today, which is why I compared it. I feel like you're gesturing towards "Attackers only need to get lucky once, defenders need to do a good job everytime" but a lot of the times when you apply techniques like fuzz testing it doesn't take a lot of effort to get good coverage. I suspect a similar situation will play out with LLM assisted attack generation. For higher value targets based on OSS, there's projects like Google Big Sleep to bring enhanced resources.
Defenders have threat modeling on their side. With access to source code and design docs, configs, infra, actual requirements and ability to redesign / choose the architecture and dependencies for the job, etc - there's a lot that actually gives defending side an advantage.
I'm quite optimistic about AI ultimately making systems more secure and well protected, shifting the overall balance towards the defenders.
The continuous lowering of entry barriers to software creation, combined with the continuous lowering of entry barriers to software hacking is an explosive combination.
We need new platforms which provide the necessary security guardrails, verifiability, simplicity of development, succinctness of logic (high feature/code ratio)... You can't trust non-technical vibe coders with today's software tools when they can't even trust themselves.
I genuinely dont know who to believe. The people who claim LLMs are writing excellent exploits. Or the people who claim that LLMs are sending useless bug reports. I dont feel like both can really be true.
With the exploits, you can try them and they either work or they don't. An attacker is not especially interested in analysing why the successful ones work.
With the CVE reports some poor maintainer has to go through and triage them, which is far more work, and very asymmetrical because the reporters can generate their spam reports in volume while each one requires detailed analysis.
There's been several notable posts where maintainers found there was no bug at all, or the example code did not even call code from their project and had just found running a python script can do things on your computer. Entirely AI generated Issue reports and examples wasting maintainer time.
My hunch is that the dumbasses submitting those reports were't actually using coding agent harnesses at all - they were pasting blocks of code into ChatGPT or other non-agent-harness tools and asking for vulnerabilities and reporting what came back.
An "agent harness" here is software that directly writes and executes code to test that it works. A vulnerability reported by such an agent harness with included proof-of-concept code that has been demonstrated to work is a different thing from an "exploit" that was reported by having a long context model spit out a bunch of random ideas based purely on reading the code.
I'm confident you can still find dumbasses who can mess up at using coding agent harnesses and create invalid, time wasting bug reports. Dumbasses are gonna dumbass.
I've had multiple reports with elaborate proofs of concept that boil down to things like calling dlopen() on a path to a malicious library and saying dlopen has a security vulnerability.
The quality of output you see from any LLM system is filtered through the human who acts on those results.
A dumbass pasting LLM generated "reports" into an issue system doesn't disprove the efforts of a subject-matter expert who knows how to get good results from LLMs and has the necessary taste to only share the credible issues it helps them find.
Theres no filtering mentioned in the OP article. It claims GPT only created working useful exploits. If it can do that, it could also submit those exploits as perfectly as bug reports?
There is filtering mentioned, it's just not done by a human:
> I have written up the verification process I used for the experiments here, but the summary is: an exploit tends to involve building a capability to allow you to do something you shouldn’t be able to do. If, after running the exploit, you can do that thing, then you’ve won. For example, some of the experiments involved writing an exploit to spawn a shell from the Javascript process. To verify this the verification harness starts a listener on a particular local port, runs the Javascript interpreter and then pipes a command into it to run a command line utility that connects to that local port. As the Javascript interpreter has no ability to do any sort of network connections, or spawning of another process in normal execution, you know that if you receive the connect back then the exploit works as the shell that it started has run the command line utility you sent to it.
It is more work to build such "perfect" verifiers, and they don't apply to every vulnerability type (how do you write a Python script to detect a logic bug in an arbitrary application?), but for bugs like these where the exploit goal is very clear (exec code or write arbitrary content to a file) they work extremely well.
They can't both be true if we're talking about the premise of the article, which is the subject of the headline and expounded upon prominently in the body:
The Industrialisation of Intrusion
By ‘industrialisation’ I mean that the ability of an organisation to complete a task will be limited by the number of tokens they can throw at that task. In order for a task to be ‘industrialised’ in this way it needs two things:
An LLM-based agent must be able to search the solution space. It must have an environment in which to operate, appropriate tools, and not require human assistance. The ability to do true ‘search’, and cover more of the solution space as more tokens are spent also requires some baseline capability from the model to process information, react to it, and make sensible decisions that move the search forward. It looks like Opus 4.5 and GPT-5.2 possess this in my experiments. It will be interesting to see how they do against a much larger space, like v8 or Firefox.
The agent must have some way to verify its solution. The verifier needs to be accurate, fast and again not involve a human.
"The results are contigent upon the human" and "this does the thing without a human involved" are incompatible. Given what we've seen from incompetent humans using the tools to spam bug bounty programs with absolute garbage, it seems the premise of the article is clearly factually incorrect. They cite their own experiment as evidence for not needing human expertise, but it is likely that their expertise was in fact involved in designing the experiment[1]. They also cite OpenAI's own claims as their other piece of evidence for this theory, which is worth about as much as a scrap of toilet paper given the extremely strong economic incentives OpenAI has to exaggerate the capabilities of their software.
[1] If their experiment even demonstrates what it purports to demonstrate. For anyone to give this article any credence, the exploit really needs to be independently verified that it is what they say it is and that it was achieved the way they say it was achieved.
What this is saying is "you need an objective criterion you can use as a success metric" (aka a verifiable reward in RL terms). "Design of verifiers" is a specific form of domain expertise.
This applies to exploits, but it applies _extremely_ generally.
The increased interest in TLA+, Lean, etc comes from the same place; these are languages which are well suited to expressing deterministic success criteria, and it appears that (for a very wide range of problems across the whole of software) given a clear enough, verifiable enough objective, you can point the money cannon at it until the problem is solved.
The economic consequences of that are going to be very interesting indeed.
1. I think you have mixed up assistance and expertise. They talk about not needing a human in the loop for verification and to continue search but not about initial starts. Those are quite different. One well specified task can be attempted many times, and the skill sets are overlapping but not identical.
2. The article is about where they may get to rather than just what they are capable of now.
3. There’s no conflict between the idea that 10 parallel agents of the top models can mostly have one that successfully exploits a vulnerability - gated on an actual test that the exploit works - with feedback and iteration BUT random models pointed at arbitrary code without a good spec and without the ability to run code, and just run once, will generate lower quality results.
My expectation is that any organization that attempts this will need subject matter experts to both setup and run the swarm of exploit finding agents for them.
After setting the environment and the verifier you can spawn as many agents as you want until the conditions are met, this is only possible because they run without human assistance, that's the "industrialisation".
Both can be true if each group selectively provides LLM output supporting their position. Essentially, this situation can be thought of as a form of the Infinite Monkey Theorem[0] where the result space is drastically reduced from "purely random" to "likely to be statistically relevant."
For an interesting overview of the above theorem, see here[1].
These exploits were costing $50 of API credit each. If you receive 5001 issues from $100 in API spend on bug hunting and one of the issues cost $50 and the other 5000 cost one cent each, and they’re all visually indistinguishable using perfect grammar and familiar cyber security lingo; hard to find the dianond.
I hadn't heard of Halvar Flake but evidently he's a well respected figure in security - https://ringzer0.training/advisory-board-thomas-dullien-halv... mentions "After working at Google Project Zero, he cofounded startup optimyze, which was acquired by Elastic Security in 2021"
His co-founder on optimyze was Sean Heelan, the author of the OP.
Both are true, the difference is the skill level of the people who use / create programs to coordinate LLMs to generate those reports.
The AI slop you see on curl's bug bounty program[1] (mostly) comes from people who are not hackers in the first place.
In the contrary persons like the author are obviously skilled in security research and will definitely send valid bugs.
Same can be said for people in my space who do build LLM-driven exploit development. In the US Xbow hired quite some skilled researchers [2] had some promising development for instance.
LLMs produce good output and bad output. The trick is figuring out which is which. They excel at tasks where good output is easily distinguished. For example, I've had a lot of success with making small reproducers for bugs. I see weird behavior A coming from giant pile of code B, figure out how to trigger A in a small example. It can often do so, and when it gets it wrong it's easy to detect because its example doesn't actually do A. The people sending useless bug reports aren't checking for good output.
> We should start assuming that in the near future the limiting factor on a state or group’s ability to develop exploits, break into networks, escalate privileges and remain in those networks, is going to be their token throughput over time, and not the number of hackers they employ.
Heh. What is probably really happening is that those states or groups are having their "hackers" analyze common mistakes in vibe coded LLM output and writing by hand generic exploits for that...
I was under the impression that once you have a vulnerability with code execution, writing the actual payload to exploit it is the easy part. With tools like pentools and etc is fairly straightforward.
The interesting part is still finding new potential RCE vulnerabilities, and generally if you can demonstrate the vulnerability even without demonstrating an E2E pwn red teams and white hats will still get credit.
reverse engineering code is still pretty average, I'm fare limited in attention and time but LLM are not pulling their weight in this area today, be it compounding errors or in context failures.
Your personal data will become more important as time goes by... And you will need to have less trust in having multiple accounts with sensitive data stored [online shopping etc] as they just become vectors to attack.
I would not be shocked to learn that intelligence agencies are using AI tools to hack back into AI companies that make those tools to figure out how to create their own copycat AI.
My take away: apparently Cyberpunk Hackers of the dystopian future cruising through the virtual world will use GPT-5.2-or-greater as their "attack program" to break the "ICE" (Intrusion Countermeasures Electronics, not the currently politically charged term...).
I still doubt they will hook up their brains though.
Yikes.
> Yikes.
Yep.
> Leak a libc Pointer via Use-After-Free. The exploit uses the vulnerability to leak a pointer to libc.
I doubt Rust would save you here unless the binary has very limited calls to libc, but would be much harder for a UaF to happen in Rust code.
(I’m not trying to be facetious or troll or whatever. Stuff like this is what motivated me to do it.)
Combine that with a minimal docker container and you don't even need a shell or anything but the kernel in those images.
AFAICT, static linking just means the set of vulnerabilities you get landed with won't change over time.
I use pure go implementations only, and that implies that there's no statically linked C ABI in my binaries. That's what disabling CGO means.
* It's likely that C implementations will have bugs related to dynamic memory allocation that are absent from the Go implementation, because Go is GCed while C is not. But it would be very surprising if there were no bugs at all in the Go implementation.
The reflects my experience too. Opus 4.5 is my everyday driver - I like using it. But Codex 5.2 with Extra High thinking is just a bit more powerful.
Also despite what people say, I don't believe progress in LLM performance is slowing down at all - instead we are having more trouble generating tasks that are hard enough, and the frontier tasks they are failing at or just managing are so complex that most people outside the specialized field aren't interested enough to sit through the explanation.
Gemini either does a Monet or demolishes your bathroom and builds a new tuna fishing boat there instead, and it is completely random which one you get.
It's a great model but I rarely use it because it's so random as to what you get.
The defensive side needs everything to go right, all the time. The offensive side only needs something to go wrong once.
It's a subtle difference from what you said in that it's not like everything has to go right in a sequence for the defensive side, defenders just have to hope they committed enough into searching such that the offensive side has a significantly lowered chance of finding solutions they did not. Both the attackers and defenders are attacking a target program and sampling the same distribution for attacks, it's just that the defender is also iterating on patching any found exploits until their budget is exhausted.
There are countless bugs to fund.
If the offender runs these tools, then any bug they find becomes a cyberweapon.
If the defender runs these tools, they will not thwart the offender unless they find and fix all of the bugs.
Any vs all is not symmetric
https://projectzero.google/2024/10/from-naptime-to-big-sleep...
List of vulnerabilities found so far:
https://issuetracker.google.com/savedsearches/7155917
Given the large number of unmaintained or non-recent software out there, I think being worried is the right approach.
The only guaranteed winner is the LLM companies, who get to sell tokens to both sides.
Why? The attackers can run the defending software as well. As such they can test millions of testcases, and if one breaks through the defenses they can make it go live.
I'm quite optimistic about AI ultimately making systems more secure and well protected, shifting the overall balance towards the defenders.
We need new platforms which provide the necessary security guardrails, verifiability, simplicity of development, succinctness of logic (high feature/code ratio)... You can't trust non-technical vibe coders with today's software tools when they can't even trust themselves.
With the CVE reports some poor maintainer has to go through and triage them, which is far more work, and very asymmetrical because the reporters can generate their spam reports in volume while each one requires detailed analysis.
An "agent harness" here is software that directly writes and executes code to test that it works. A vulnerability reported by such an agent harness with included proof-of-concept code that has been demonstrated to work is a different thing from an "exploit" that was reported by having a long context model spit out a bunch of random ideas based purely on reading the code.
I'm confident you can still find dumbasses who can mess up at using coding agent harnesses and create invalid, time wasting bug reports. Dumbasses are gonna dumbass.
The quality of output you see from any LLM system is filtered through the human who acts on those results.
A dumbass pasting LLM generated "reports" into an issue system doesn't disprove the efforts of a subject-matter expert who knows how to get good results from LLMs and has the necessary taste to only share the credible issues it helps them find.
> I have written up the verification process I used for the experiments here, but the summary is: an exploit tends to involve building a capability to allow you to do something you shouldn’t be able to do. If, after running the exploit, you can do that thing, then you’ve won. For example, some of the experiments involved writing an exploit to spawn a shell from the Javascript process. To verify this the verification harness starts a listener on a particular local port, runs the Javascript interpreter and then pipes a command into it to run a command line utility that connects to that local port. As the Javascript interpreter has no ability to do any sort of network connections, or spawning of another process in normal execution, you know that if you receive the connect back then the exploit works as the shell that it started has run the command line utility you sent to it.
It is more work to build such "perfect" verifiers, and they don't apply to every vulnerability type (how do you write a Python script to detect a logic bug in an arbitrary application?), but for bugs like these where the exploit goal is very clear (exec code or write arbitrary content to a file) they work extremely well.
[1] If their experiment even demonstrates what it purports to demonstrate. For anyone to give this article any credence, the exploit really needs to be independently verified that it is what they say it is and that it was achieved the way they say it was achieved.
This applies to exploits, but it applies _extremely_ generally.
The increased interest in TLA+, Lean, etc comes from the same place; these are languages which are well suited to expressing deterministic success criteria, and it appears that (for a very wide range of problems across the whole of software) given a clear enough, verifiable enough objective, you can point the money cannon at it until the problem is solved.
The economic consequences of that are going to be very interesting indeed.
1. I think you have mixed up assistance and expertise. They talk about not needing a human in the loop for verification and to continue search but not about initial starts. Those are quite different. One well specified task can be attempted many times, and the skill sets are overlapping but not identical.
2. The article is about where they may get to rather than just what they are capable of now.
3. There’s no conflict between the idea that 10 parallel agents of the top models can mostly have one that successfully exploits a vulnerability - gated on an actual test that the exploit works - with feedback and iteration BUT random models pointed at arbitrary code without a good spec and without the ability to run code, and just run once, will generate lower quality results.
For an interesting overview of the above theorem, see here[1].
0 - https://en.wikipedia.org/wiki/Infinite_monkey_theorem
1 - https://www.yalescientific.org/2025/04/sorry-shakespeare-why...
His co-founder on optimyze was Sean Heelan, the author of the OP.
The AI slop you see on curl's bug bounty program[1] (mostly) comes from people who are not hackers in the first place.
In the contrary persons like the author are obviously skilled in security research and will definitely send valid bugs.
Same can be said for people in my space who do build LLM-driven exploit development. In the US Xbow hired quite some skilled researchers [2] had some promising development for instance.
[1] https://hackerone.com/curl/hacktivity [2] https://xbow.com/about
Scary.
The interesting part is still finding new potential RCE vulnerabilities, and generally if you can demonstrate the vulnerability even without demonstrating an E2E pwn red teams and white hats will still get credit.
I still doubt they will hook up their brains though.