Hey HN, Michael here, co-founder of Promptfoo and one of the people working on the Codex Security CLI at OpenAI.
Thanks for checking this out and for flagging the auth issues. We just open-sourced it, and there's still plenty for us to improve. Expect the product to evolve quickly.
If you try it, I'd really appreciate hearing what works well and what you think we should improve. Happy to answer questions here.
How does it deal with the current guardrails 5.6 Sol has on finding vulnerabilities? When I use it in the Codex app it would sometimes say it found a vulnerability, but it cannot tell me what it is.
Thanks! You've run into a real limitation: the CLI doesn't bypass the model's cybersecurity guardrails. If GPT-5.6 Sol finds a vulnerability but refuses to explain it, switching from the Codex app to the CLI won't automatically fix that.
For authorized defensive work, Trusted Access for Cyber (TAC1/Daybreak) can reduce refusals depending on the model and the account or organization where access is provisioned. It isn't a blanket bypass.
If you're an open-source maintainer, you can apply for conditional Codex Security access here:
If you have an example of "found a vulnerability but won't tell me what it is," I'd love to take a look too. You can send it to use with /feedback (or message me).
Or try Grok, 4.5 seems pretty capable, should be close to K3 in many coding tasks. I use it for code review of what other "stronger" models shit out (like Sol) and it constantly finds even pretty big bugs or just not robust enough solutions (Sol tends to overengineer, yes, but I'm not so sure it overengineers the right parts, so far my experience woth it has been mid. Except it understanding my drawings and collages and it being capable of far better frontend/design dev than 5.4 or even 5.5 was).
Sounds like that's the only solution. I'm so sick of this safety nonsense I was going to switch from Anthropic to OpenAI because of it. I'm so disappointed to see it's just more of the same.
Model finds a vulnerability in your code but "refuses" to tell you. Words can hardly express the sheer absurdity of it.
> GPT-5.6 Sol finds a vulnerability but refuses to explain it
I think it would be a good practice to refund the session cost in that case. Otherwise a customer just spent some money in order to get exactly nothing.
I tried it, it started a scan but stopped after hitting the rate-limit of my account. It gave up after just a minute of retrying (rate limits are tokens per minute, so... :P).
It said "Partial output was kept at <...>", but I dont see a obvious way of picking it up in a new scan? (The failed run cost me ~$13)
Yeah, you're right. A per-minute rate limit shouldn't kill a scan after a minute, and "partial output was kept" makes it sound like you can pick up where you left off. You can't yet, unfortunately. --max-cost can limit estimated spend, but we still need proper retries and resume. Sorry you spent $13 finding that out. Please send me an email and I'll help make it right.
> Have experience shipping production full-stack products across modern web frontends and backend services.
I'm amazed that the requirements are so low (or at least this vague) for jobs at companies like these.
Has anyone else had the experience of going to an interview and feeling like you were never asked any qualifying questions?
All the questions were easy, your answers were straightforward, you "got them right", but then were not chosen?
I find on the other side, they're also left with dozens of people who "passed" and then it comes down to a pretty arbitrary decision on who gets hired (if we are talking external, no referral, etc.)
I wonder if they can make job descriptions highly specific to filter the shortlist faster and more effectively (to actually get a shortlist).
I interviewed with them and had a great experience despite it ultimately not being a good fit due to communication issues and a bad flu coming on in the final interview - they, and Michael especially, was very kind and understanding.
That's quite an indictment of the common app development practices. (Not that I disagree with your point...)
And security is hard. Because it is by definition off the happy path, it is quite often at odds with MVPs and rapid release cycles. Then you add all the ways the users can use your product to attack/abuse others.
Any non-hobbyist app development does indeed require at least a decent understanding of security.
Credit card entry, I think you should know a few basics like don't put it in MongoDB?? Or nah? It's just like any other user data?
How about a background check - can anyone take a user-entered DL and randomly Google stuff to see what they find?
Can I store your SSN in plain text in a text file? Why not?
The user had to upload their ID for IDV but I use Vercel. I guess I have to put it on S3. What should the bucket policy be for all these driver's license photos - there are so many???
> Credit card entry, I think you should know a few basics like don't put it in MongoDB?? Or nah? It's just like any other user data?
> How about a background check - can anyone take a user-entered DL and randomly Google stuff to see what they find?
> Can I store your SSN in plain text in a text file? Why not?
You wouldn't be touching any of those unless you work for a handful of providers where that's their whole business. Usually you add the dependency, use their widget, and that's it.
If you're handling credit card numbers yourself, you're in a shrinking subset of developer roles. I've spent the last 7 years of my career working at payment processors, so I do handle that stuff, but the majority of my industry has built an infrastructure that makes it so most developers don't have to think about that.
To your other examples, not everyone on the team needs to know these things up front. Someone in the review process does, and eventually that knowledge gets disseminated and more people know it to carry it forward in their career.
Not having clear, objective criteria enables arbitrary decisions (not against OP, I mean in general.. and in general I dislike this pattern a lot). On the extreme other end of the spectrum would be 100% objective criteria, and companies being forced to pick a random applicant that matches them. If they want only the best, they have to have high expectations, but be able to actually define them. You say "cultural fit", I say "corruption", let's blow this whole joint.
> You say "cultural fit", I say "corruption", let's blow this whole joint.
It's not necessarily corruption. If there's no conflict between principal and agent, it's fine.
Just like when I sent my butler to go and buy a bottle of wine, he can make arbitrary choices, but that doesn't mean he's corrupt or going against my wishes. I trust his judgement, and since it's a repeated game our incentives are aligned.
All fair points and I guess I would dislike it more if it hadn’t benefitted me more. In general, if I can get face-to-face with a human then I have a huge advantage, and any barriers to getting to that point are a net negative to me.
At the start of my career I was under-credentialed, and had to rely on lax requirements to get myself in front of people. I might never have gotten anywhere if my first few employers hadn’t been willing to overlook a lack of degree or commercial (rather than open-source) experience, directly as a result of a funnel with a wide entrance.
It’s not corruption to simply hire the people you subjectively feel a preference for working for, instead of objective criteria. It’s not public tax funds that fuel salaries, it’s your own money. You get to spend it how you like.
Yes, many jurisdictions have outlawed arbitrary discrimination against protected classes (eg race), which is an entirely different matter, and not what we are discussing here.
Yeah, and it's not necessarily mobbing to only tell people you really like about your party. But where it occurs, people never admit it to themselves and justify it in such a way, so that justification is meaningless. All wars of aggression are called a defensive emergency measure. Hundreds and thousands and rarely would anyone say "we'll take this because we can and you're helpless". No matter how glaringly obvious it is, it's never admitted.
In the same way, I can acccept "cultural fit" as a summary of things a person can describe, sure. But I think more often than not it's just a thought-terminating cliché. It can also just mean "I cannot verbalize my reasons and/or don't want to admit to them".
You can say if something fits only if you either can describe both sides in sufficient detail and where it wouldn't fit, e.g. a plug and a socket. But if it's dark, you barely see anything, and just have a "hunch", then "fit" doesn't even apply. It's like telling someone you won't let them through a door because they wouldn't "fit" anyway -- okay, so let them try, if they actually won't fit you don't need to read tea leaves and gate keep based on that.
Preferring to go with a more safe and familiar and obvious candidate, fine. But don't pretend it's because the others won't "fit".
Culture, in so far as it deserves the name, shapes the people exposed to or in it, as well as the other way around. E.g. if only people who fit the culture can work at a company, no company can exist in the first place, because for there to be a culture there need to be people there. So that leaves setting the culture in stone after it grew to a certain size, and only looking for more of the same, which also isn't great, but at least still honest.
If a culture is so brittle it cannot integrate people who aren't already a product of it, that may be a legitimate choice of the company, but my assessment to find that lame is also valid.
I feel the same way about immigration troubles, tangentially. We moan because people we don't actively try to get to know don't care for our rules which we don't enforce in a confident, but respectful manner. We basically require sterile, bland input because we have no immune system worth speaking of and no way to process and refine what comes in.
I agree the phrase "culture fit" is weasel-wordy, but...
> You can say if something fits only if you either can describe both sides in sufficient detail and where it wouldn't fit
Many human dynamics, including sexual attraction, love, and even just who will be fun or easy to work with, are dynamics we don't fully understand, and cannot fully specify. In all these cases "I'll know when it see it" is perfectly reasonable, and need not be hiding an untoward motive. Which is not say, ofc, that it can't be hiding such a motive. That happens too.
It depends whether your goal is to hire for specific knowledge (hence specific questions) or for overall mindset and abilities (hence broader questions where you are able to extract the way a person thinks)
If you didn't already know, jobs at highly competitive companies tend to have vague job requirements because they expect to be able to apply your raw intelligence to changing demands quickly. There's no point being hyper-specific about the exact software packages because that's not what they want. What they want is someone who, after talking to an interviewer for 30 minutes, leaves them with the thought "Wow, this person can do anything we need of them. They can probably tell us what we need too and take ownership of large projects. Hire!"
Agree. Being smart, competent, and high in conscientiousness is more important than any highly specific “qualification”. It’s not about checking a bunch of boxes. If you have a track record of getting shit done, you’ll have something to contribute.
> There's no point being hyper-specific about the exact software packages because that's not what they want
Okay. This makes it sound like they're more sophisticated but it seems more like they are less sophisticated, less specific, and a lot more vague in the job descriptions they themselves create.
If you look at any technical role, game dev or something where people are building important things at scale - there are a lot of specifics. Libraries, methodologies, where if you didn't know them you are nowhere near a fit.
I'm just wondering. It's OpenAI. Surely there is some domain-specific something beyond "has experience shipping front-end and back-end services" since that includes basically everyone.
I see where you're coming from, but as someone who regularly interviews engineers, I don't care about specific tech stacks when evaluating a candidate very much either. I can only think of two positions I've worked in where such a thing really mattered.
A good engineer can adapt and catch up without a lot of lead time. For a contractor, I'd be much more specific - but for someone who's going to join my team? I'm looking for a candidate that can demonstrate their problem solving ability, creative thinking and communication skills.
Other than having some kind of experience in the general domain we work in, those "soft skills" are far harder to find than specific tech experience.
Not every engineering job is entry level or as simple as most fullstack crud. Deeper into industry you find highly specific well defined positions for a given domain. Soft skills matter more the higher the ladder but id take a killer senior who can be difficult over a team of mediocre staff engineers.
I agree, I recently interviewed at Synthiolabs, and they asked me two questions: one about RAG and two about graph RAG. They rejected me even though I answered correctly, and the interviewer was also a college kid.
The plugin, including when invoked through the Codex CLI, is great for scanning the repo you're currently working in. The standalone Security CLI/SDK uses the same scanner, but is built for running security across many repos over time: org-wide scans, historical results, deduplication, false-positive tracking, budget controls, and CI integration.
We've been talking to hundreds of engineering and security teams, and their feedback is shaping what we build.
Like Promptfoo, our goal is practical tooling that fits into the workflows teams already have.
We are actively working on officially supporting this. Because it's open source it is pretty easy to point a coding agent at it now and switch out the model.
Not yet, unfortunately. We only just opened the repo, and there isn't a public issue specifically tracking local or OpenAI-compatible endpoint support.
> co-founder of Promptfoo and one of the people working on the Codex Security CLI at OpenAI.
> Thanks for checking this out and for flagging the auth issues.
Offtopic, but this right here is why I don't believe any marketing around "great amazing models that one-shot everything and programmers are no longer needed".
You just have to look at what these labs routinely produce, and their own products.
Edit to respond to @simonw whose comment I saw before he retracted it ;)
This comment is tied directly to consistent continuous claims by the LLM labs. Their own products disprove their own claims, and it would indeed be nice if fewer people believed them :)
Hey looks cool. I tried to run this on a small oss library and here's what happened:
$ codex-security scan .
[00:00] Preparing scan
[00:00] Authentication: stored Codex credentials.
[00:01] Preparing scan
[00:42] Running scan
[00:42] Preflight: worker delegation supported (up to 8 worker slots).
[41:03] Running scan
codex-security: This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber
codex-security: Partial output was kept at /Users/ryan/.codex/state/plugins/codex-security/scans/framework/codex-security-framework-z7eNfr.
Just some feedback, but it ran for over 40 minutes and during that time I had no idea what was happening, thought it was frozen or in a bad state. Also, it ate through 25% of my weekly credits :(
Same happened to me. Very disappointing, this should be mentioned before the user even authenticates in the onboarding phase of the product asking the user if they have acquired the whitelisting from OpenAI and want to proceed or not! But running for 35+ minutes plus to give such a response is very disappointing and very awkward! Not to mention the lost weekly tokens!
I was going to switch to OpenAI and away from Anthropic because of "safety" nonsense like this. Really disappointed to discover it's just gonna be more of the same. Looks like Chinese models are the only ones without any of this safety bullshit.
Just ran it on a small repo. It ran for almost an hour and then got interrupted. It drained half my weekly usage on a Pro plan.
npx codex-security scan .
[00:00] Preparing scan
[00:00] Authentication: stored Codex credentials.
[00:03] Preparing scan
[01:20] Running scan
[01:20] Preflight: worker delegation supported (up to 8 worker slots).
[52:47] Running scan
codex-security: Could not save the Codex Security scan: Repository HEAD changed while the scan was running. Start a new scan.
codex-security: Partial output was kept at ...
They just opened the weights. I expect competition from various providers will drop the price a bit. But you're right. I'd hope a model closer to GLM 5.2's price would be sufficiently useful.
A license and presumably revshare is required for large scale inference as a service of Kimi K3. There's clearly something going on right now with every router at the same or higher price as the Moonshot list price.
So I wouldn't count on competition if $/mil token is actually set by Moonshot, but I would expect $/tok to drop when there's a more competitive frontier open weight model.
> If the Licensee or any of its affiliates operates a Model as a Service business,
and the aggregate revenue of the Licensee and its affiliates exceeds 20 million
US dollars (or the equivalent in other currencies) in total over any consecutive
12 months, the Licensee must enter into a separate agreement with Moonshot AI
before using the Software or its derivative works for any commercial purpose.
Only if "the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months"
Oof, that's a bad outcome. Half your weekly usage and a 50-minute scan just to get a HEAD error at the end is not acceptable. --max-cost can help limit estimated spend, but that doesn't fix the underlying problem or give you your quota back. We need to handle a changing checkout and partial results much better. Sorry you ran into this. Please send me an email.
security tools from AI companies feel like fire departments run by arsonists. useful, sure, but you can't help noticing who benefits from all the fires
That’s exactly what they’re saying. With the added (and very important) detail that the people selling the fireproofing are the the same who armed everyone with flamethrowers. Why shouldn’t someone complain about that?
I've been noticing that many new projects that would have been written in Python or Node a year ago are starting to be written in Go, Rust, etc.
Theory: people realized there’s little benefit to Python for agents. As Zep wrote, an “agent is a long-running, concurrent, I/O-bound process that spends most of its time waiting on a model, a tool, or a human[1]” — not a particular strength of Python.
I'm wondering if you'd considered Go (or others—Go’s just my fav ) before landing on Node, and more broadly whether you've noticed a similar pattern?
> As Zep wrote, an “agent is a long-running, concurrent, I/O-bound process that spends most of its time waiting on a model, a tool, or a human[1]” — not a particular strength of Python.
That sounds exactly like a strength of Python, no? Python is excellent at working IO blocks and waiting in general being interpreted language with first-class async support.
Right. Python is excellent at waiting on things instead of actually doing things, and on the off case it does do things, most of the time it's really juggling strings around the thing instead of doing it.
s/
I generally don't write Python, but like others, I disagree with GP too. In fact, a lot of my work involves Python being written now, simply because that's what LLMs like to write.
I suspect a lot of the Python and TypeScript code is getting prompted by folks who don't know any difference between dynamically typed and statically typed.
In the early days of LLM coding, it seemed to be much better at Python for whatever reason. I was never a big Python guy, but I got better results so I ran with it. That definitely doesn't seem to be the case anymore. Last week I asked Codex to mash up Super Mario Bros and Contra ROMs and it just did everything in straight assembly and absolutely crushed it. It couldn't do that 2 years ago. Python is just momentum and I think it's going to die down now that the models are much better.
I think it's because python is far more approachable/ubiquitous than go/rust. It's the entry level language for many people from all disciplines of life. Scientific community uses it, data science uses it.
Golang/rust however are very convenient to distribute. Small, portable, fast exe's are very nice. With agentic coding golang/rust are now accessible to a lot more people.
I see where you're coming from, but to be honest I respectfully disagree. If all you mean to do is writing small one-off scripts then sure Python may be the right tool[1], but when you're doing something more complicated Go is just simpler. And for an LLM, Go is even better for all the reasons mentioned in sibling comments and OP.
Yes, now that humans write less than 99% of code, the most important criteria for a language isn't readability, which I'd argue was always Python's main selling point, but the underlying runtime. There are practical limits to how fast a Python program can run either under I/O or CPU bound compared to other popular and mature languages with extensive libraries, like Elixir, Go or C++, depending on your use case.
So you're fine with your code having unintended behaviour, as long as that unintended behaviour passes a test that the agent wrote?
My preferred approach is to read and understand everything the LLM produces AND have it create test suites (which I also read and understand). The LLM can help you with that too - just have it breakdown and explain the code at each iteration.
If it's code for a pacemaker, it should be reviewed by multiple humans. If it's code for an unimportant side project, I will never read it. Between those extremes are shades of grey.
Part of our job in this new era is to understand the worst-case consequences of a bug given how the code interacts with the world, then allocate our effort based on that understanding. This can only be done on a case by case basis.
Not really. I test the output thoroughly, I examine the thinking process, I go through the diff to see if anything jumps out but I my thinking process/the way I work has changed. Low level programming thinking has gotten atrophied it seems.
Update: As far as I understand, this was already available as a Codex plugin. The main news is that OpenAI has now open-sourced it, and development is still moving quickly.
Some of approaches there could be useful in other contexts. OAI has the compute to experiment with different prompts and I'd expect these to be somewhat optimized.
Yes, I think this is an under-appreciated part of the release. I hope people can adapt them to their own workflows. We run A LOT of evals as the Promptfoo team and we've spent billions of tokens fine-tuning them. You can expect more skills as we branch out to other security workflows and further improvements to the codex security prompts.
Not often I see companies referring to HN, thanks “ We quietly released the open-source Codex Security CLI, but Hacker News found it before we had a chance to share it here…”
codex-security scan .
[00:00] Preparing scan
[00:00] Authentication: stored Codex credentials.
[00:01] Preparing scan
[31:25] Running scan
codex-security: This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber
I mean... wasn't this the intended goal? And you wasted all my tokens for nothing?!
I seem to have gotten a bunch of you are trying to stuff we don't allow errors.. very annoying.
Can they explain what types of projects it works on and how does it check I own it? Like will it just not work on Linux kernel even on my own patches to it?
Fair question, and I agree the refusals are frustrating.
The CLI doesn't do a repository-ownership check. Public projects are supported, and reviewing your own Linux kernel patches is the kind of defensive work we want to support.
The refusals come from model guardrails, which can be overly cautious. Trusted Access for Cyber (TAC1/Daybreak) is a separate, approved access path that can reduce those refusals.
I don't think I should share it publicly since it was a proprietary piece of code but is there a way to not have it waste so many tokens if it fails this feels very very maddening seeing your tokens burn but get zilch for it in return maybe I and my company(in API costs it burnt over 100+$ of tokens a good chunk of my weekly limit for nothing) are too poor for it...
Getting into the Cyber program seems like a hassle as a freelance/open source person with tiny projects. Think used in production at 2-3 companies but only 20 something stars(ofc I don't market it but it just feels very unfair).
Really glad to see this open-sourced. One thing I'd be interested in is how you think about the balance between false positives and false negatives. In practice, developers tend to stop trusting security tools if they generate too much noise, but missing a real issue is obviously costly too.
In short, this isn't an offline scanner. The CLI runs locally but the code and context needed for analysis are sent to the hosted model (OpenAI).
For API, Business, and Enterprise accounts, business data isn't used to train models by default. Retention and other data controls depend on the product and account configuration.
If your company doesn't allow source code to leave its environment, you shouldn't run this against that codebase. Local and third-party endpoints aren't officially supported yet, but you can read through the code and your favorite coding agent will allow you to use it with any model of your choice in 30 seconds.
Anything you ever do with any non-locally-hosted model always "uploads code" to the inference provider because that's how it works: the model uses tools to inspect the code, the result of the tool use is sent in an API call to provide context (and a prompt for the next turn), and then the response continues the process.
This is true and has to be true for any hosted model that works with existing code: it's not specific to this application.
Yes, if only due to the business models being entirely different. AWS sells compute. OpenAI sells models. One of these things benefits from training on your data significantly more than the other.
I don't understand why Snyk is IN business in any way. Who really wants to upload his own code to a company that is specialized at searching security issues?
How can I trust that they show me all findings they have instead of selling the best ones to some three letter organisations?
One of their sales people made fun of me via email. Apparently they believe that not being their customers means you cannot possibly know if a dependency you use has an active CVE.
Also they haven't figured out codeberg exists, so the resume page of a project of mine on snyk[1] still links to github and reports the project as "inactive", having the last commit 2 years ago, and the last release 2 months ago. I think it's quite telling of their quality.
Looks great but the CLI output is not particularly interesting while the scan is running. I wish it could show token usage, some kind of progress, etc.
I was actually discussing solutions for this with my coworkers—building white-hat security agents. It seems like openai/codex-security could simplify a lot of that, or at least provide a version of Codex that's purpose-built for security workflows. Really exciting news!
One thing worth checking in the meantime: OPENAI_API_KEY or CODEX_API_KEY can override an existing ChatGPT/Codex login. If you're trying to use your ChatGPT login, run this in bash or zsh:
unset OPENAI_API_KEY CODEX_API_KEY
Then retry your scan.
If it still fails, could you share the exact error and whether you're using ChatGPT login or an API key? Happy to help debug. You can also file an issue in the repo and we'll take a look!
By default, you can sign in with your ChatGPT/Codex account or use an OPENAI_API_KEY. It also does not require cyber registration but it can help if you encounter refusals. If you give it a try, please feel free to message me, I would love your feedback.
Thing is, you WILL encounter refusals with Sol doing anything remotely adjacent to security work. Which for Codex Security is kinda... problematic.
Just a few days back, I was reviewing some small bit of legacy DSA signature verification code, to get a sense of how safe it is to reuse - purely defensive, precautionary work and the context of it was there. But I simply wasn't able to use Codex Security: it threw refusal tantrums on every step of the way. Even the reasoning went like "nah, this is false positive, this is defensive code hardening, I'll nuke the subagent and tell it so" , followed by a refusal.
In the end, I was only able to do partial review with vanilla Codex w/o Codex Security.
The scanner is the least interesting part of this. The harness around it is the product: dedup across runs, false-positive tracking, budget controls, CI gating. That is the layer where we'll see most interesting innovations in my opinion.
I'm building AQ, a coding harness for teams and the pattern is identical. For a while, I thought the raw model is the answer and quickly changed my mind. Purpose built harnesses are way more powerful than it sounds.
"If we had only used this tool, OpenAI would've had to pay off another patsy for their marketing stunt" -- Huggingface
"The 'S' in OpenAI is for "Security". Ever since we developed this tool we've had almost zero AI generated reports of outbreaks of allegedly-rogue AI agents breaking out of allegedly-secure testing environments, probably" -- OpenAI
Thanks for checking this out and for flagging the auth issues. We just open-sourced it, and there's still plenty for us to improve. Expect the product to evolve quickly.
If you try it, I'd really appreciate hearing what works well and what you think we should improve. Happy to answer questions here.
CLI docs: https://learn.chatgpt.com/docs/security/cli
EDIT: If you'd like to help make this better, we're hiring: https://openai.com/careers/full-stack-software-engineer-cybe...
How does it deal with the current guardrails 5.6 Sol has on finding vulnerabilities? When I use it in the Codex app it would sometimes say it found a vulnerability, but it cannot tell me what it is.
For authorized defensive work, Trusted Access for Cyber (TAC1/Daybreak) can reduce refusals depending on the model and the account or organization where access is provisioned. It isn't a blanket bypass.
If you're an open-source maintainer, you can apply for conditional Codex Security access here:
https://openai.com/form/codex-for-oss/
For enterprise teams, the public Daybreak onboarding guide is here:
https://help.openai.com/en/articles/20001261-enterprise-dayb...
If you have an example of "found a vulnerability but won't tell me what it is," I'd love to take a look too. You can send it to use with /feedback (or message me).
Or perhaps a better option is to use something like Kimi K3 and cancel the GPT subscription altogether.
Model finds a vulnerability in your code but "refuses" to tell you. Words can hardly express the sheer absurdity of it.
> https://openai.com/form/codex-for-oss/
Hey, Lead maintainer of vim here. Applied twice already never heard anything back. This is a frustrating experience!
I think it would be a good practice to refund the session cost in that case. Otherwise a customer just spent some money in order to get exactly nothing.
It said "Partial output was kept at <...>", but I dont see a obvious way of picking it up in a new scan? (The failed run cost me ~$13)
I'm amazed that the requirements are so low (or at least this vague) for jobs at companies like these.
Has anyone else had the experience of going to an interview and feeling like you were never asked any qualifying questions?
All the questions were easy, your answers were straightforward, you "got them right", but then were not chosen?
I find on the other side, they're also left with dozens of people who "passed" and then it comes down to a pretty arbitrary decision on who gets hired (if we are talking external, no referral, etc.)
I wonder if they can make job descriptions highly specific to filter the shortlist faster and more effectively (to actually get a shortlist).
Anyway end rant. Cool job, hope you fill it.
Maybe this is part of the problem
And security is hard. Because it is by definition off the happy path, it is quite often at odds with MVPs and rapid release cycles. Then you add all the ways the users can use your product to attack/abuse others.
Any non-hobbyist app development does indeed require at least a decent understanding of security.
It is extremely rare for companies to roll their own payment processing anymore, or even handle PCI scope at all.
How about a background check - can anyone take a user-entered DL and randomly Google stuff to see what they find?
Can I store your SSN in plain text in a text file? Why not?
The user had to upload their ID for IDV but I use Vercel. I guess I have to put it on S3. What should the bucket policy be for all these driver's license photos - there are so many???
> How about a background check - can anyone take a user-entered DL and randomly Google stuff to see what they find?
> Can I store your SSN in plain text in a text file? Why not?
You wouldn't be touching any of those unless you work for a handful of providers where that's their whole business. Usually you add the dependency, use their widget, and that's it.
If you're handling credit card numbers yourself, you're in a shrinking subset of developer roles. I've spent the last 7 years of my career working at payment processors, so I do handle that stuff, but the majority of my industry has built an infrastructure that makes it so most developers don't have to think about that.
To your other examples, not everyone on the team needs to know these things up front. Someone in the review process does, and eventually that knowledge gets disseminated and more people know it to carry it forward in their career.
(indeed that first wildcard means any account)
It's not necessarily corruption. If there's no conflict between principal and agent, it's fine.
Just like when I sent my butler to go and buy a bottle of wine, he can make arbitrary choices, but that doesn't mean he's corrupt or going against my wishes. I trust his judgement, and since it's a repeated game our incentives are aligned.
At the start of my career I was under-credentialed, and had to rely on lax requirements to get myself in front of people. I might never have gotten anywhere if my first few employers hadn’t been willing to overlook a lack of degree or commercial (rather than open-source) experience, directly as a result of a funnel with a wide entrance.
Yes, many jurisdictions have outlawed arbitrary discrimination against protected classes (eg race), which is an entirely different matter, and not what we are discussing here.
In the same way, I can acccept "cultural fit" as a summary of things a person can describe, sure. But I think more often than not it's just a thought-terminating cliché. It can also just mean "I cannot verbalize my reasons and/or don't want to admit to them".
You can say if something fits only if you either can describe both sides in sufficient detail and where it wouldn't fit, e.g. a plug and a socket. But if it's dark, you barely see anything, and just have a "hunch", then "fit" doesn't even apply. It's like telling someone you won't let them through a door because they wouldn't "fit" anyway -- okay, so let them try, if they actually won't fit you don't need to read tea leaves and gate keep based on that.
Preferring to go with a more safe and familiar and obvious candidate, fine. But don't pretend it's because the others won't "fit".
Culture, in so far as it deserves the name, shapes the people exposed to or in it, as well as the other way around. E.g. if only people who fit the culture can work at a company, no company can exist in the first place, because for there to be a culture there need to be people there. So that leaves setting the culture in stone after it grew to a certain size, and only looking for more of the same, which also isn't great, but at least still honest.
If a culture is so brittle it cannot integrate people who aren't already a product of it, that may be a legitimate choice of the company, but my assessment to find that lame is also valid.
I feel the same way about immigration troubles, tangentially. We moan because people we don't actively try to get to know don't care for our rules which we don't enforce in a confident, but respectful manner. We basically require sterile, bland input because we have no immune system worth speaking of and no way to process and refine what comes in.
> You can say if something fits only if you either can describe both sides in sufficient detail and where it wouldn't fit
Many human dynamics, including sexual attraction, love, and even just who will be fun or easy to work with, are dynamics we don't fully understand, and cannot fully specify. In all these cases "I'll know when it see it" is perfectly reasonable, and need not be hiding an untoward motive. Which is not say, ofc, that it can't be hiding such a motive. That happens too.
Okay. This makes it sound like they're more sophisticated but it seems more like they are less sophisticated, less specific, and a lot more vague in the job descriptions they themselves create.
If you look at any technical role, game dev or something where people are building important things at scale - there are a lot of specifics. Libraries, methodologies, where if you didn't know them you are nowhere near a fit.
I'm just wondering. It's OpenAI. Surely there is some domain-specific something beyond "has experience shipping front-end and back-end services" since that includes basically everyone.
It makes this job look like a Starbucks role.
A good engineer can adapt and catch up without a lot of lead time. For a contractor, I'd be much more specific - but for someone who's going to join my team? I'm looking for a candidate that can demonstrate their problem solving ability, creative thinking and communication skills.
Other than having some kind of experience in the general domain we work in, those "soft skills" are far harder to find than specific tech experience.
Not every engineering job is entry level or as simple as most fullstack crud. Deeper into industry you find highly specific well defined positions for a given domain. Soft skills matter more the higher the ladder but id take a killer senior who can be difficult over a team of mediocre staff engineers.
We've been talking to hundreds of engineering and security teams, and their feedback is shaping what we build.
Like Promptfoo, our goal is practical tooling that fits into the workflows teams already have.
Really glad we got to ship this, and there's still a lot we want to improve in Codex Security and in Promptfoo!
The issue tracker is here: https://github.com/openai/codex-security/issues
If you open an issue with the endpoint or model you want to use, I'd be happy to follow up there.
> Thanks for checking this out and for flagging the auth issues.
Offtopic, but this right here is why I don't believe any marketing around "great amazing models that one-shot everything and programmers are no longer needed".
You just have to look at what these labs routinely produce, and their own products.
Edit to respond to @simonw whose comment I saw before he retracted it ;)
This comment is tied directly to consistent continuous claims by the LLM labs. Their own products disprove their own claims, and it would indeed be nice if fewer people believed them :)
Promotions all around.
For context can you share the line count?
So I wouldn't count on competition if $/mil token is actually set by Moonshot, but I would expect $/tok to drop when there's a more competitive frontier open weight model.
The current price is likely a result of the high demand and the high requirements of this model.
> If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.
Looks as if these companies could wait until they reach $20 million of revenue with Kimi K3 until they enter a separate agreement.
If ya wouldn't mind crediting me a quick 60 months that would be great.
Can’t speak to the results, but the cost isn’t high.
I've been noticing that many new projects that would have been written in Python or Node a year ago are starting to be written in Go, Rust, etc.
Theory: people realized there’s little benefit to Python for agents. As Zep wrote, an “agent is a long-running, concurrent, I/O-bound process that spends most of its time waiting on a model, a tool, or a human[1]” — not a particular strength of Python.
I'm wondering if you'd considered Go (or others—Go’s just my fav ) before landing on Node, and more broadly whether you've noticed a similar pattern?
1: https://blog.getzep.com/agentic-development-in-go/
That sounds exactly like a strength of Python, no? Python is excellent at working IO blocks and waiting in general being interpreted language with first-class async support.
s/
I generally don't write Python, but like others, I disagree with GP too. In fact, a lot of my work involves Python being written now, simply because that's what LLMs like to write.
I can’t fathom why anybody would want to continue working with dynamically typed languages when they can now get types for free.
Did this create a new game? And did you publish the result? Might’ve misunderstood.
Golang/rust however are very convenient to distribute. Small, portable, fast exe's are very nice. With agentic coding golang/rust are now accessible to a lot more people.
I see where you're coming from, but to be honest I respectfully disagree. If all you mean to do is writing small one-off scripts then sure Python may be the right tool[1], but when you're doing something more complicated Go is just simpler. And for an LLM, Go is even better for all the reasons mentioned in sibling comments and OP.
[1]: I tend to rely more on Bash, though...
please tell me you're reading the AI code
My preferred approach is to read and understand everything the LLM produces AND have it create test suites (which I also read and understand). The LLM can help you with that too - just have it breakdown and explain the code at each iteration.
Part of our job in this new era is to understand the worst-case consequences of a bug given how the code interacts with the world, then allocate our effort based on that understanding. This can only be done on a case by case basis.
Some of approaches there could be useful in other contexts. OAI has the compute to experiment with different prompts and I'd expect these to be somewhat optimized.
https://x.com/openai/status/2082263717916586117?s=46&t=mnfnj...
I’m confused. Why are you thanking them for that?
codex-security scan . [00:00] Preparing scan [00:00] Authentication: stored Codex credentials. [00:01] Preparing scan [31:25] Running scan codex-security: This content was flagged for possible cybersecurity risk. If this seems wrong, try rephrasing your request. To get authorized for security work, join the Trusted Access for Cyber program: https://chatgpt.com/cyber
I mean... wasn't this the intended goal? And you wasted all my tokens for nothing?!
Can they explain what types of projects it works on and how does it check I own it? Like will it just not work on Linux kernel even on my own patches to it?
The CLI doesn't do a repository-ownership check. Public projects are supported, and reviewing your own Linux kernel patches is the kind of defensive work we want to support.
The refusals come from model guardrails, which can be overly cautious. Trusted Access for Cyber (TAC1/Daybreak) is a separate, approved access path that can reduce those refusals.
If you're an open-source maintainer, you can apply for conditional Codex Security access here: https://openai.com/form/codex-for-oss/
For enterprise teams, the Daybreak onboarding process is explained here: https://help.openai.com/en/articles/20001261-enterprise-dayb...
If you have a specific repro, I'd be happy to look into it.
Getting into the Cyber program seems like a hassle as a freelance/open source person with tiny projects. Think used in production at 2-3 companies but only 20 something stars(ofc I don't market it but it just feels very unfair).
https://github.com/alibaba/open-code-review
For API, Business, and Enterprise accounts, business data isn't used to train models by default. Retention and other data controls depend on the product and account configuration.
If your company doesn't allow source code to leave its environment, you shouldn't run this against that codebase. Local and third-party endpoints aren't officially supported yet, but you can read through the code and your favorite coding agent will allow you to use it with any model of your choice in 30 seconds.
More on OpenAI's enterprise data handling: https://openai.com/enterprise-privacy/
This is true and has to be true for any hosted model that works with existing code: it's not specific to this application.
How can I trust that they show me all findings they have instead of selling the best ones to some three letter organisations?
Also they haven't figured out codeberg exists, so the resume page of a project of mine on snyk[1] still links to github and reports the project as "inactive", having the last commit 2 years ago, and the last release 2 months ago. I think it's quite telling of their quality.
1. https://security.snyk.io/package/pip/typedload
https://github.com/openai/codex-security/pull/22
One thing worth checking in the meantime: OPENAI_API_KEY or CODEX_API_KEY can override an existing ChatGPT/Codex login. If you're trying to use your ChatGPT login, run this in bash or zsh:
Then retry your scan.If it still fails, could you share the exact error and whether you're using ChatGPT login or an API key? Happy to help debug. You can also file an issue in the repo and we'll take a look!
Edit: there's a little bit more meat here: https://github.com/openai/codex-security/tree/main/sdk/types...
Just a few days back, I was reviewing some small bit of legacy DSA signature verification code, to get a sense of how safe it is to reuse - purely defensive, precautionary work and the context of it was there. But I simply wasn't able to use Codex Security: it threw refusal tantrums on every step of the way. Even the reasoning went like "nah, this is false positive, this is defensive code hardening, I'll nuke the subagent and tell it so" , followed by a refusal.
In the end, I was only able to do partial review with vanilla Codex w/o Codex Security.
https://wasnotwas.com/writing/inside-openai-codex-security/
https://pornoscan.us/codex-security-analysis.html
I'm building AQ, a coding harness for teams and the pattern is identical. For a while, I thought the raw model is the answer and quickly changed my mind. Purpose built harnesses are way more powerful than it sounds.
"If we had only used this tool, OpenAI would've had to pay off another patsy for their marketing stunt" -- Huggingface
"The 'S' in OpenAI is for "Security". Ever since we developed this tool we've had almost zero AI generated reports of outbreaks of allegedly-rogue AI agents breaking out of allegedly-secure testing environments, probably" -- OpenAI