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.
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.
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?
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.
We have found a large number of vulnerabilities in major open-source projects at HunterSwarm[0], and we're disclosing them on a typically 30-day timeline, but we're working with the projects. I personally think 90 days is way too long.
(With customers, however, we integrate right into their CI-CD, but because of the nature of our work, we can only work with certain customers. Major open source projects, however, have a lot of stakeholders and typically no budget, so we just try to help them as much as we can.)
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!”
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.
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.
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?
Starting today, our policy is “Full 90”, i.e. vulnerabilities will be disclosed 90 days after they are reported to the vendor.
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.
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.
If AI is doing all the work I could see that; same day remediation.
Which means developers must trust AI blindly since there is no time to review the fix. Which also means that developers are completely unneeded
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.
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 the people finding the vulns aren't the people running the services.
Why aren't they? The people running the services can use the AI to find the vulnerabilities before they release.
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.
Is the AI in the room with you right now?
Unrelated, but didn't realise .google was a TLD.
We have found a large number of vulnerabilities in major open-source projects at HunterSwarm[0], and we're disclosing them on a typically 30-day timeline, but we're working with the projects. I personally think 90 days is way too long.
(With customers, however, we integrate right into their CI-CD, but because of the nature of our work, we can only work with certain customers. Major open source projects, however, have a lot of stakeholders and typically no budget, so we just try to help them as much as we can.)
0. https://swarmresearch.org
Nice unsolicited plug for your slop product.
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!”
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.
Project Zero has been around far longer than GenAI or LLMs, so your point is not really relevant here.
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.
You say that like it's a bad thing...
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?