Don't write or fix the code for me (thanks but I can manage that on my own with much less hassle), but instead tell me which places in the code look suspicious and where I need to have a closer look.
When I ask Claude to find bugs in my 20kloc C library it more or less just splits the file(s) into smaller chunks and greps for specific code patterns and in the end just gives me a list of my own FIXME comments (lol), which tbh is quite underwhelming - a simple bash script could do that too.
ChatGPT is even less useful since it basically just spend a lot of time to tell me 'everything looking great yay good job high-five!'.
So far, traditional static code analysis has been much more helpful in finding actual bugs, but static analysis being clean doesn't mean there are no logic bugs, and this is exactly where LLMs should be able to shine.
If getting more useful potential-bugs-information from LLMs requires an extensively customized setup then the whole idea is getting much less useful - it's a similar situation to how static code analysis isn't used if it requires extensive setup or manual build-system integration instead of just being a button or menu item in the IDE or enabled by default for each build.
The writing has been on the wall with so called hallucinations where LLMs just make stuff up that the hype was way out over its skiis. The examples of lawyers being fined for unchecked LLM outputs being presented as fact type of stories will continue to take the shine off and hopefully some of the raw gungho nature will slow down a bit.
Here's a technique that often works well for me: When you get unexpectedly poor results, ask the LLM what it thinks an effective prompt would look like, e.g. "How would you prompt Claude Code to create a plan to effectively review code for logic bugs, ignoring things like FIXME and TODO comments?"
The resulting prompt is too long to quote, but you can see the raw result here: https://gist.github.com/CharlesWiltgen/ef21b97fd4ffc2f08560f...
From there, you can make any needed improvements, turn it into an agent, etc.
I asked ChatGPT to analyze its weaknesses and give me a pre-prompt to best help mitigate them and it gave me this: https://pastebin.com/raw/yU87FCKp
I've found it very useful to avoid sycophancy and increase skepticism / precision in the replies it gives me
It still needs guidance, but it quashed bugs yesterday that I've previously spent many days on without finding a solution for.
It can be tricky, but they definitely can be significant aid for even very complex bugs.
Definitely optimistic for this way to use AI
Where I still need to extend this, is to introduce function calling in the flow, when "it has doubts" during reasoning, would be the right time to call out a tool that would expand the context its working with (pull in other files, etc).
Yeah, don't listen to "wisdom of the crowd" when it comes to LLM models, there seems to be a ton of fud going on, especially on subreddits.
GPT-OSS was piled on for being dumb in the first week of release, yet none of the software properly supported it at launch. As soon as it was working properly in llama.cpp, it was clear how strong the model was, but at that point the popular sentiments seems to have spread and solidified.
I explicitly asked it to read all the code (within Cline) and it did so, gave me a dozen action items by the end of it, on a Django project. Most were a bit nitpicky, but two or three issues were more serious. I found it pretty useful!
I get it though, non programmers or weak programmers don't scrutinise the results and are more likely to be happy to pay. Still, bit of a shame.
Maybe these tools exist, but at least to me, they don't surface among all the noise.
Even very simple prompts can yield very useful outputs.
“Report each bug you spot in this code with a markdown formatted report.” worked better than I expected.
It costs just a couple of dollars to scan through an entire codebase with something like Gemini Flash.
I often use Claude/GPT-5/etc to analyze existing repositories while deliberately omitting the tests and documentation folders because I don't want them to influence the answers I'm getting about the code - because if I'm asking a question it's likely the documentation has failed to answer it already!
but i've been limiting it to a lot less than 20k LoC, i'm sticking with stuff i can just paste into the chat window.
GPT 5 has been disappointing with thinking and without.
Some history: https://hn.algolia.com/?q=curl+AI
I'll be doing a retrospective in a few weeks when the dust has settled, as well as new tools I've been made aware of.
Seems like ZeroPath might be worth looking into if the price is reasonable
This is notable given Daniel Stenberg's reports of being bombarded by total slop AI-generated false security issues in the past: https://www.linkedin.com/posts/danielstenberg_hackerone-curl...
Concerning HackerOne: "We now ban every reporter INSTANTLY who submits reports we deem AI slop. A threshold has been reached. We are effectively being DDoSed. If we could, we would charge them for this waste of our time"
Also this from January 2024: https://daniel.haxx.se/blog/2024/01/02/the-i-in-llm-stands-f...
A few of these PRs are dependabot PRs which match on "sarif", I am guessing because the string shows up somewhere in the project's dependency list. "Joshua sarif data" returns a more specific set of closed PRs. https://github.com/curl/curl/pulls?q=is%3Apr+Joshua+sarif+da...
It's primarily from people just throwing source code at an LLM, asking it to find a vulnerability, and reporting it as-read, without having any actual understanding of if it is or isn't a vulnerability.
The difference in this particular case is it's someone who is: 1) Using tools specifically designed for security audits and investigations. 2) Takes the time to read and understand the vulnerability reported, and verifies that it is actually a vulnerability before reporting.
Point 2 is the most significant bar that people are woefully failing to meet and wasting a terrific amount of his time. The one that got shared from a couple of weeks ago https://hackerone.com/reports/3340109 didn't even call curl. It was straight up hallucination.
https://joshua.hu/llm-engineer-review-sast-security-ai-tools... ("Hacking with AI SASTs: An overview of 'AI Security Engineers' / 'LLM Security Scanners' for Penetration Testers and Security Teams")
I guess mastodon link is simply a confirmation that bugs were indeed bugs, even with wrong code snippets?
Tools included ZeroPath, Corgea and Almanax.
You can read about my experience here: https://codepathfinder.dev/blog/introducing-secureflow-cli-t...
Old post: https://shivasurya.me/security-reviews/sast/2024/06/27/autom...
The set seems to be:
https://joshua.hu/llm-engineer-review-sast-security-ai-tools...
So he likes ZeroPath. Does that get us any further? No, the regular subscription costs $200 and the free one-time version looks extremely limited and requires yet another login.
Also of course, all low hanging fruit that these tools detect will be found quickly in open source (provided that someone can afford a subscription), similar to the fact that oss-fuzz has diminishing returns.
You can see the fixes that resulted from this in the PRs that mention "sarif" in the curl repository: https://github.com/curl/curl/pulls?q=is%3Apr+sarif+is%3Aclos...
It’s like a police facial recognition, they can help police but there is no way they are “replacing police”
* Clearly useful to people who are already competent developers and security researchers
* Utterly useless to people who have no clue what they're doing
But the latter group's incompetency does not make AI useless in the same way that a fighter jet is not useless because a toddler cannot pilot it.
Imagine what your doctors will be like two generations down the road.
> Utterly useless to people who have no clue what they're doing
> the same way that a fighter jet is not useless
AI is currently like a bicycle, while we were all running hills before.
There's a skill barrier and getting less complicated each week.
The marketing goal is to say "Push the pedal and it goes!" like it was a car on a highway, but it is a bicycle, you have to keep pedaling.
The effect on the skilled-in-something-else folks is where this is making a difference.
If you were thinking of running, the goal was to strengthen your tendons to handle the pavement. And a 2hr marathon pace is almost impossible to do.
Like a bicycle makes a <2hr marathon distance "easy" for someone who does competitive rowing, while remaining impossible for those who have been training to do foot races forever.
Because the bicycle moves the problem from unsprung weights and energy recovery into a VO2 max problem, also into a novel aerodynamics problem.
And if you need to walk a rock garden, now you need to lug the bike too with you. It is not without its costs.
This AI thing is a bicycle for the mind, but a lot of people go only downhill and with no brakes.
I’m a reasonable developer with 30+ years of experience. Recently I worked on an API design project and had to generate a mock implementation based on a full openapi spec. Exactly what Copilot would be good at. No amount of prompting could make it generate a fully functional spring-boot project doing both the mock api and present the spec at a url at the same time. Yet it did a very neat job at just the mock for a simpler version of the same api a few weeks prior. Go figure.
https://news.ycombinator.com/item?id=38845878
https://news.ycombinator.com/item?id=43907376
https://media.ccc.de/v/froscon2025-3407-ai_slop_attacks_on_t...
I disagree.
I'm making a board game of 6 colors of hexes, and I wanted to be able to easily edit the board. The first time around, I used a screenshot of a bunch of hexagons and used paint to color them (tedious, ugly, not transparent, poor quality). This time, I asked ChatGPT to make an SVG of the board and then make a JS script so that clicking on a hex could cycle through the colors. Easier, way higher quality, adjustable size, clean, transparent.
It would've taken me hours to learn and set that up for myself, but ChatGPT did it in 10min with some back and forth. I've made one SVG in my life before this, and never written any DOM-based JS scripts.
Yes, it's a toy example, but you don't have to knwo what you're doing to get useful things from AI.
You might be underestimating the expertise you applied in these 10 minutes. I know I often do.
> it's a toy example
This technology does exceptionally well on toy examples, I think because there are much fewer constraints on acceptable output than ‘real’ examples.
> you don't have to knwo what you're doing to get useful things from AI
You do need to know what is useful though, which can be a surprisingly high bar.
You're someone who knows the difference between a PNG and an SVG, knows enough Javascript to know that "DOM-based" JS is a thing, and has presumably previously worked in software/IT.
You're smart enough to know things, and you're also smart enough to know there's a lot that you don't know.
That's a far cry from the way a lot of laypeople, college kids, and fully nontechnical people try to use LLMs.
Sounds like it was a lot more than 22, assuming most are valid.
Somethings we learnt alone the way, is that when it comes to specifically this field of security what we called low-level security (memory security etc.), validation and debugging had became more important than vulnerability discovery itself because of hallucinations.
From our trial-and-errors (trying validator architecture, security research methodology e.g., reverse taint propagation), it seems like the only way out of this problem is through designing a LLM-native interactive environment for LLMs, validate their findings of themselves through interactions of the environment or the component. The reason why web security oriented companies like XBOW are doing very well, is because how easy it is to validate. I seen XBOW's LLM trace at Black Hat this year, all the tools they used and pretty much need is curl. For web security, abstraction of backend is limited to a certain level that you send a request, it whether works or you easily know why it didn't (XSS, SQLi, IDOR). But for low-level security (memory security), the entropy of dealing with UAF, OOBs is at another level. There are certain things that you just can't tell by looking at the source but need you to look at a particular program state (heap allocation (which depends on glibc version), stack structure, register states...), and this ReACT'ing process with debuggers to construct a PoC/Exploit is what been a pain-in-the-ass. (LLMs and tool callings are specifically bad at these strategic stateful task, see Deepmind's Tree-of thoughts paper discussing this issue) The way I've seen Google Project Zero & Deepmind's Big Sleep mitigating this is through GDB scripts, but that's limited to a certain complexity of program state.
When I was working on our integration with GGML, spending around two weeks on context, tool engineering can already lead us to very impressive findings (OOBs); but that problem of hallucination scales more and more with how many "runs" of our agentic framework; because we're monitoring on llama.cpp's main branch commits, every commits will trigger a internal multi-agent run on our end and each usually takes around 1 hours and hundreds of agent recursions. Sometime at the end of the day we would have 30 really really convincing and in-depth reports on OOBs, UAFs. But because how costly to just validate one (from understanding to debugging, PoC writing...) and hallucinations, (and it is really expensive for each run) we had to stop the project for a bit and focus solving the agentic validation problem first.
I think when the environment gets more and more complex, interactions with the environment, and learning from these interactions will matters more and more.
Thanks for sharing your experience ! It correlates with this recent interview with Sutton [1]. That real intelligence is learning from feedback with a complex and ever changing environment. What an LLM does is to train on a snapshot of what has been said about that environment and operate on only on that snapshot.
Red borders around every slide and very flashy images
https://joshua.hu/llm-engineer-review-sast-security-ai-tools...
https://mastodon.social/@icing@chaos.social/1152440641434357...
>tldr
>The code was correct, the naming was wrong.
The key word is "potential", though. They're still wildly unpredictable and unreliable, which is why an expert human is required to validate their output.
The big problem is the people overhyping the technology, selling it as "AI", and the millions deluded by the marketing. Amidst the false advertising, uncertainty, and confusion, people are forced to speculate about the positive and negative impacts, with wild claims at both extremes. As usual, the reality is somewhere in the middle.
Many people advocate for the use of AI technology for SAST testing. There are even people and companies that deliver SAST scanners based on AI technology. However: Most are just far from good enough.
In the best case scenario, you’ll only be disappointed. But the risk of a false sense of security is enormous.
Some strong arguments against AI scanners can be found on https://nocomplexity.com/ai-sast-scanners/
They're using it correctly. It's a system of tools, not an autopilot.
If you read Corgea's (one of the products used) "whitepaper", it seems that AI is not the main show:
> BLAST addresses this problem by using its AI engine to filter out irrelevant findings based on the context of the application.
It seems that AI is being used to post-process the findings of traditional analyzers. It reduces the amount of false positives, increasing the yield quality of the more traditional analyzers that were actually used in the scan.
Zeropath seems to use similar wording like "AI-Enabled Triage" and expressions like "combining Large Language Models with AST analysis". It also highlights that it achieves less false positives.
I would expect someone who developed this kind of thing to setup a feedback loop in which the AI output is somehow used to improve the static analysis tool (writing new rules, tweaking existing ones, ...). It seems like the logical next step. This might be going on on these products as well (lots of in-house rule extensions for more traditional static analysis tools, written or discovered with help of AI, hence the "build with AI" headline in some of them).
Don't get me wrong, this is cool. Getting an AI to triage a verbose static analysis report makes sense. However, it does not mean that AI found the bugs. In this model, the capabilities of finding relevant stuff are still capped at the static analyzer tools.
I wonder if we need to pay for it. I mean, now that I know it is possible (at least in my head), it seems tempting to get open source tools, set them to max verbosity, and find which prompts they are using on (likely vanilla) coding models to get them to triage the stuff.
We do not use traditional static analyzers; our engine was built from the ground up to use LLMs as a primitive. The issues ZeroPath identified in Joshua's post were indeed surfaced and triaged by AI.
If you're interested in how it works under the hood, some of the techniques are outlined here: https://zeropath.com/blog/how-zeropath-works
Joshua describes it as follows: "ZeroPath takes these rules, and applies (or at least the debug output indicates as such) the rules to every .. function in the codebase. It then uses LLM’s ability to reason about whether the issue is real or not."
Would you say that is a fair assessment of the LLM role in the solution?
That's an editorialized headline (so it may get fixed by dang and co) - if you click through to what Daniel Stenberg said he was more clear:
> Joshua Rogers sent us a massive list of potential issues in #curl that he found using his set of AI assisted tools.
AI-assisted tools seems right to me here.
Also, think about it: of course I read Joshua's report. Otherwise, how could I have known the names of the products he used?
How would you have worded it?
Even Joshua's blog post does not clearly state which parts and how much is "AI". Neither does the pdf.
You did this with an AI and you do not understand what you're doing here: https://news.ycombinator.com/item?id=45330378
That makes its results unpredictable.
So don’t have AI create your bugs.
Instead have your AI look for problems - then have it create deterministic tools and let tools catch the issues in a repeatable, understandable, auditable way. Have it build short, easy to understand scripts you can commit to your repo, with files and line numbers and zero/nonzero exit codes.
It’s that key step of transforming AI insights into detection tools that transforms your outcomes from probabilistic to deterministic. Ask it to optimize the tools so they run in seconds. You can leave them in the codebase forever as linters, integrate them in your CI, and never have that same bug again.
