I thought they made a strong case for removing the additional 30 day patch rollout time. But reducing the entire process to 30 days total seems needlessly harsh and I didn't think it was justified by the article at all. What if the bug is actually inherently hard to fix?
lol, and if its in a widespread library in an embedded device? Thanks for compromising the internet, aka there is a tension inherent, not that i disagree with policy and the forcing function underneath, but it is quite a different perspective for non corporate projects.
Crowdstrike was a good example of why you should never allow global updates to roll out all at once , even for configuration changes.
Cloudflare has learnt the same lesson recently https://blog.cloudflare.com/18-november-2025-outage/
Bug fix speed isn’t particularly relevant, rollout speed is.
Every piece of software you have ever used almost certainly has critical security vulnerabilities, right now, no matter how up to date it is. However, you don't know about them and neither does anyone else.
Obviously you can't prove that, but again, you can't prove that for literally any software you're using right now.
Let's say a bug hasn't been found for the past 20 years. Google finds it, but it's going to take longer than 30 days to fix. Is it really better to let everyone know about it after 30 days? If no one else has found it (true, as far as we know) in the past 20 years, what are the probabilities it will be found by a second person in the next few months?
Perhaps eventually, they can finally live up to their namesake. On January 1 2029, we will move to a "Full 0" policy where vulnerabilities are disclosed after (looks at camera) zero days.
IMO the disclosure should ~day-two after there is a fix/workaround available
Releasing before that is non-actionable by many actors (aside from taking the service completely down I guess). Any time after fix is available is just allowing vulnerable systems to be attacked and is playing favourites with who gets fix and who doesn't
No, that's not what it means. It just means that the benefit must outweigh the risk, as always. Making surgical patches to existing code is also quite different to introducing features, which is what developers typically spend their time on. Not that same-day patching would imply literally zero time to review either.
That does not follow. Somebody stills needs to set it up. And maybe to sign off. The AI could instantly diagnose and summarize the problem and its remediation and you say the word.
The way you describe it puts the role of humans as merely decorarive and ceremonial. The work is done by AI and actually validated by AI too, and increasingly the role of the human is to give an ok to whatever the AI outputs
No, they're not - they can just release and wait for another vulnerability disclosure. Then the loop between [Vendors] and Project Zero needs no humans in it.
If AI is doing all the work then why would remediation be needed at all? Why can't the software AI produces be vuln-free if the same AI is used to detect them?
Because LLMs are not deterministic or predictable and make mistakes that they are capable of fixing.
LLMs will write vulnerable code while the same model can find those vulnerabilities if you prompt it right. Some stats from a previous post was that the CVEs found in Linux were costing on average $1000 in api usage each. So having the LLM prevent writing vulnerabilities would explode the costs.
You talk as though there is one AI model that is perfect and capable of identifying all potential flaws with no mistakes or omissions, and that access is free and unlimited.
Project zero is run by Google, one of the few groups doing frontier models. If you consider their revenue streams, Google is the sole frontier player that can actually fund insane research and compute out of pocket without relying on any hype cycle.
So, Google will have some model that is much more powerful than the publicly available models. Before releasing it to the world, they can use it to mitigate any issue it discovers anywhere
Exactly. That's super irresponsible to knowingly release software that you haven't used AI to check if it had any vulns. There's zero (heh) excuse to release vulnerable software at that point.
Catching lots of bugs in software written with poor methodology is the new norm and that is not the same as releasing software that is not vulnerable. That is releasing software that needs increasing technological improvements or luck to
find remaining vulnerabilities.
Running trivy scans on my docker containers has really highlighted which projects keep up to date. I’ve been moving away from all the one person projects to more commercial offerings that keep on top of updates.
Of course I don’t expect a single person to work every day for free, but I also can’t have a project with only a single dev exposed to the internet.
It makes to hold out for disclosure if there's a pay day involved. Since this is Google, the bug bounty is not the issue. I say release right away. Nothing else will compel the company being reported on to fix otherwise.
I’m not familiar with this specific effort but it seems like the companies digging around for the bugs should also offer similar access to their products that discovered the vulnerabilities to the vendors or OSS projects to patch them. Otherwise vendors might use this as a leveraged approach to get people to adopt LLMs, specifically theirs, so their products remain reasonably attractive. Google, OpenAI, Anthropic etc could just turn on existing software as a strategy to get more adoption of their tools.
“Don’t like our constant stream of patches? For $10k/mo + tokens you can get access to the model to patch the vulnerabilities we disclose!”
How is it irrelevant? Just because an effort existed before some change, doesn’t mean it can be repurposed and leveraged for other purposes. It’s the entire reason law and policy for government isn’t set in stone and is adaptable, because culture, technology and so on change.
The dynamic in this context absolutely has changed. Marginal cost to discover vulnerabilities from AI frontier labs and the big players has dropped significantly. The cost to patch is low, if you have access to certain resources. If you don’t, tough luck.
So this argument is essentially against the entire argument of AI alignment guardrails right now. Maybe you disagree with those, but I for one am happy they’re doing some things to protect against bio related warfare engineering. This is a case of cyber that’s about grey. Even there the labs have limited access to some and are choosing winners.
For sure I would love to get access to P0 researchers for a m 10k/months, but i don’t believe you can hire such talent for this sum.
They are one of or maybe the most skilled security research team that’s doing public research and their blog is the source of many modern exploitation techniques
For a while, asking your agent to act like a P0 researcher was a nice nudge to find better bugs !
Usually you can’t find the kind of bugs P0 finds by just throwing AI at it. You need highly skilled people driving/guiding it. Highly skilled people are the bottleneck.
In any case it’d be sensible to make yourself familiar with the “specific effort” before you make up your opinion on it.
That's kinda the problem with asinine "safety" idea the AI moguls have.
It is just leaving people with less safe software. At least claude the limitations are actively preventing you from looking for security bugs in your own software
I mean, sure, the cited $10k / mo is exploitative for small open-source projects but... in general, why wouldn't we want there to be pressure on companies and projects to use high-quality LLM-based checks if they want to see wide adoption?
Is this a version of extortion? Isn't this Google, a massive company with lots of power and control, forcing the prioritization of work on everyone under the guise of security? It's at least clearly a threat.
Project Zero doesn't create the vulnerabilities, it finds them. Vendors want unlimited time to fix their bugs. Users want to know ASAP if the software they are running is vulnerable. These things are obviously in tension, so you need a deadline.
It's gotten easier finding vulnerabilities, so every day without disclosure now brings a higher chance that blackhats have already found the bug. The deadline gets shorter accordingly.
Calling vulnerability research "extortion" is the kind of ignorant messenger-blaming the hacker culture has always ridiculed.
This is sort of what I was getting at with my original post. I'm just curious as to the power dynamics here. Not everyone has $0.5 trillion in revenue every year.
On February 2, 2027, we will move to a “Full 60” policy where vulnerabilities are disclosed after 60 days.
On January 3, 2028, we will move to a “Full 30” policy where vulnerabilities are disclosed after 30 days.
I thought they made a strong case for removing the additional 30 day patch rollout time. But reducing the entire process to 30 days total seems needlessly harsh and I didn't think it was justified by the article at all. What if the bug is actually inherently hard to fix?
The article says they are willing to negotiate in exceptional cases, and mentions Spectre.
The testing can take a while, assuming it's a consideration at all.
Skipping testing is how you end up with another Crowdstrike scenario.
Bug fix speed isn’t particularly relevant, rollout speed is.
Let's say a bug hasn't been found for the past 20 years. Google finds it, but it's going to take longer than 30 days to fix. Is it really better to let everyone know about it after 30 days? If no one else has found it (true, as far as we know) in the past 20 years, what are the probabilities it will be found by a second person in the next few months?
Releasing before that is non-actionable by many actors (aside from taking the service completely down I guess). Any time after fix is available is just allowing vulnerable systems to be attacked and is playing favourites with who gets fix and who doesn't
2 days after there's a fix, with a max of 90 days after disclosure?
Because you still need to have the 90 days no matter what to induce the vendor to create a fix. Otherwise (many) just sit on it forever.
You’re replying in a thread whose top level comment is that the 90 days will rapidly decrease to 30.
Hope 30 is enough to induce them to fix.
Maybe the future fixes will be “Reddit + Claude”, done at home.
It’s agents all the way down.
LLMs will write vulnerable code while the same model can find those vulnerabilities if you prompt it right. Some stats from a previous post was that the CVEs found in Linux were costing on average $1000 in api usage each. So having the LLM prevent writing vulnerabilities would explode the costs.
Reality is more complicated than that.
So, Google will have some model that is much more powerful than the publicly available models. Before releasing it to the world, they can use it to mitigate any issue it discovers anywhere
Definitely much happier. There’s nothing I miss from Ubuntu.
Of course I don’t expect a single person to work every day for free, but I also can’t have a project with only a single dev exposed to the internet.
https://en.wikipedia.org/wiki/List_of_Internet_top-level_dom...
They are all owned by Google, but registrations are not open to the public.
“Don’t like our constant stream of patches? For $10k/mo + tokens you can get access to the model to patch the vulnerabilities we disclose!”
The dynamic in this context absolutely has changed. Marginal cost to discover vulnerabilities from AI frontier labs and the big players has dropped significantly. The cost to patch is low, if you have access to certain resources. If you don’t, tough luck.
So this argument is essentially against the entire argument of AI alignment guardrails right now. Maybe you disagree with those, but I for one am happy they’re doing some things to protect against bio related warfare engineering. This is a case of cyber that’s about grey. Even there the labs have limited access to some and are choosing winners.
They are one of or maybe the most skilled security research team that’s doing public research and their blog is the source of many modern exploitation techniques
For a while, asking your agent to act like a P0 researcher was a nice nudge to find better bugs !
In any case it’d be sensible to make yourself familiar with the “specific effort” before you make up your opinion on it.
It is just leaving people with less safe software. At least claude the limitations are actively preventing you from looking for security bugs in your own software
I mean, sure, the cited $10k / mo is exploitative for small open-source projects but... in general, why wouldn't we want there to be pressure on companies and projects to use high-quality LLM-based checks if they want to see wide adoption?
It's gotten easier finding vulnerabilities, so every day without disclosure now brings a higher chance that blackhats have already found the bug. The deadline gets shorter accordingly.
Calling vulnerability research "extortion" is the kind of ignorant messenger-blaming the hacker culture has always ridiculed.
What types of vulnerabilities does Project Zero find, in terms of their severity? And who are the companies that they are notifying?
You have to admit that there is likely a power imbalance here.