Artificial Intelligence 4 min read 647 words

Project Zero shortens disclosure deadlines because AI no longer gives anyone a break

ES
Project Zero shortens disclosure deadlines because AI no longer gives anyone a break

When someone finds a security bug in a program, the usual thing is to tell the vendor privately and give them some time to fix it before going public. That time is an unwritten pact: you keep quiet, they fix it, and when the patch is out everything gets published. Project Zero, the Google team that has been hunting bugs in other people’s software since 2014, has announced it’s changing the rules of that pact, and the reason is none other than AI.

From “90 + 30” to “Full X”

Since 2021, Project Zero applied the “90 + 30” policy: the vendor had up to 90 days to fix the bug and, if they fixed it, another 30 days for users to install the patch before details were published. With quite a few exceptions: grace periods, variants of the same bug, checks on whether it was fixed or not…

The new policy is much simpler, they call it “Full X”: the bug is published X days after notifying the vendor, fixed or not, and that X keeps going down:

FromDeadline
November 13, 202690 days
February 2, 202760 days
January 3, 202830 days

Almost all exceptions disappear. Only three remain: bugs already being exploited (which follow a different Google policy), those someone else has already published, and those the vendor itself asks to publish earlier. Nor will the severity the vendor assigns to a bug matter, because that doesn’t change the likelihood of an attacker finding it.

Why

The post explains it plainly, and almost everything comes down to AI:

  • Finding bugs is faster, and so is exploiting them. That makes it much more likely that several people, not all of them well-intentioned, find the same bug at the same time. Keeping the secret no longer protects much, because the secret may not be one.
  • The patch gives the bug away. Comparing the patched version with the previous one to work out what was fixed has always been possible, but now it’s almost instant. According to Project Zero, the time from patch release to someone being able to exploit the bug can now be considered zero, so those extra 30 days gave a false sense of security.
  • Volume. They’ve gone from reporting a few bugs a year to several a week, and negotiating case by case with each vendor whether something is fixed has become unmanageable.

What they ask of vendors is to fix faster, reduce the time between a patch being released and users installing it, and prepare for bugs being published soon, with plans that don’t rely only on the patch (for example, being able to temporarily disable the affected code).

The other side

This connects with something we saw a few days ago when talking about curl turning 30: VulnCheck’s tracking of AI-found bugs in Anthropic’s Project Glasswing counted almost 38,000 recorded findings and only 217 fixed, and in curl’s case, 19 of the 34 published reports weren’t vulnerabilities. AI finds bugs at a pace nobody had seen before, but whoever has to fix them is still a person, often a single person maintaining a project in their spare time.

For a large vendor with a dedicated security team, 30 days is demanding but manageable. For an open source project with a couple of maintainers, it can be very little time, especially if reports keep pouring in from other sources too (some good, some not). Project Zero’s post doesn’t go into that difference.

Even so, I find the reasoning hard to argue with. If attackers have AI too, a long deadline doesn’t protect the user, it only buys the vendor time. And the user, who is the one exposed, is the one who needs protecting. What used to be a gentlemen’s agreement between whoever finds the bug and whoever fixes it now has a third party, AI, that hasn’t signed anything.