Quo Vadis, CVD?
Coordinated vulnerability disclosure with advisories and mitigation advice has been the state of the art in cyber security for a long time. Is this era coming to an end?
The process codified in CVD-policies (e.g., the Austrian one ) and laws like the Cyber Resilience Act is based on the assumption that is possible to talk about vulnerabilities and give advice on how to mitigate exposure, without disclosing the details of the problem, let alone giving away the methods how to exploit the vulnerability.
Of course, there always were discussions about what can be shared. For example, Microsoft had a program where they shared details about upcoming patches with commercial security firms, so that their detection engines were ready once the release dropped. As far as I know, that was shut down because of leaks.
Reverse-engineering the vulnerability from the fixed software is not new at all. Malicious Threat Actors have been doing it for ages, and benevolent security researchers have been treating this a bit like a game. The write-ups of the WatchTowr analysts are usually a fun read and give good insights into that process.
Large multi-party CVD was always hard, really hard. Looking back to the Kaminsky DNS bug or the Heartbleed disclosure process shows that it was the rare exception that these things went smoothly. There are always people who think that they should get priority access and someone almost always leaks.
But now the game changed fundamentally.
We have seen more and more evidence that moden LLMs can be used to quickly pinpoint vulnerabilities and develop exploits based on the source code and the content of a typical advisory. This is especially true if mitigation advice is included.
We’ve seen this with PHP2Shell: A colleague in the EU CSIRTs Network reported that he plugged the PHP source code and the proposed mitigations (e.g., session.upload_progress.enabled = Off) in a local LLM, and got a pointer to the crucial place in the code within minutes.
The recent disclosure on XWEB's time-based root password mentioned "a constant nonce hidden in the code". Again, getting the algorithm + secret with the help of an LLM is massively easier than purely manual reverse engineering.
Or look at this thread on BSky by @pwnallthethings.bsky.social:
you don't even need the patch! I've seen several times now where someone just alluding to some general properties of the bug is sufficient for the model to hone in on it, find it, and build a working proof of concept, and it can do that in less than an hour
All this meant that the old “we have a week to apply the patch” approach got under increasing pressure. There is hardly any safe window for patching left. A good data-source for hard numbers is zerodayclock.com .
Google’s Project Zero just announced that it will significantly compress its timelines.
Multiple strategies have evolved to counter this, including:
- Silent patching: Just sneak the security fix into an otherwise harmless software update and pray that most users simply update before anybody takes a closer look at what was changed.
- Mix it with a rewrite: Tools for binary diffing are pretty advanced these days, thus a simple new “if (some condition) abort” is a clear indication that the attacker should investigate how to create exactly this condition. But if there is a major rewrite of the whole function, then it might not be obvious what the relevant change was.
- Staggered rollout: One company has started to distribute security updates to (some of) its customers under NDA. A few weeks later, the advisory is published and the update is now open to everyone. The hope here is that the NDA holds and nobody is reversing during that time span.
- Full Disclosure: Others are arguing that CVD failed and we’re better off if we just go for instant full disclosure to level the playing field.
What does this mean for our work?
The disclosure process of the PHP2Shell vulnerability a few weeks ago is a good example. Originally planned for Thursday, the release was moved up a day because the mitigation advice had leaked and thus the first PoC code was also available earlier than anticipated.
We got some information pretty early in the process and decided to stick to the TLP markings and not to communicate the upcoming release date to our constituency. But someone, somewhere did and the Internet did what it always does: once the information avalanche starts, it cannot be stopped.
Not everybody was happy with our approach. The Austrian CERT association met last week (one day before it turned 15 years old) and we had a long discussion on this topic.
The consensus was that we should communicate whatever we can to make a quick reaction by the other teams possible. This should include:
- Which team will be needed? Operating System, Webserver, Application development, Windows, Linux, Firewalling, Networking, …
- When will the announcement happen?
I didn’t expect that it would take just a week before we needed put this new policy into practice for the first time.
The feedback from the colleagues from other CSIRTs was positive – they knew why we are doing this and they are used to getting early glimpses of possible danger from us.
And then we did something new: we also gave the same heads-up message to a broader constituency. That raised many eyebrows and was not universally appreciated.
On the other hand, they are all pretty much used to the same thing if it comes to Microsoft’s patch day.
The tricky part for us is how to gauge whether an announcement we hear of is worth the pre-notification to larger constituency. We often don’t have the details; thus, we cannot independently confirm the severity. Even with details, the prediction which vulnerability will explode, and which will be a dud is often just a flip of a coin.
And then there is the preparedness paradox: Good defense can make a problem look like it didn't need that response.
So maybe it will become more or less routine in the future: The release of important patches will not be preceded by mitigation advice, just the announcement of when something happens and who it will concern.