I am tired
I am sure I am not the only in this situation. There are more Open Source maintainers and defenders who are in the same situation. From last July I started poking into various libraries/projects/deployments of open source projects to find vulnerabilites, and it became very tiring 16+ hours long days. As of today morning I found 90+ such issues, many are in projects written or used by Sunet, many has other bigger userbase too.
The things I am poking at are boring Infrastructure pieces. "National Research and Education Network/NREN" (s) in every country are not the top receivers of grant money. Small number of people working years after years to keep the infrastructure on.
And not many are that lucky to have access to the modern LLMs or knowldge on how to use them to find and fix issues. I am lucky and thankful to Anthropic and OpenAI that they provided me with access to latest models with capabilities to do this work. But, not many has it, this crates too much pressure in my mind, and for sure friends told me similar stories (in other fields). To defend Sweden (or your respective country) we will need more capabilites and compute/hardware and people willing to do this work. Where we are not dependent on another country's political views to defend our own country.
Many will say open weight models will solve this. If you only think about capabilities, yes they are catching up. Funniest part for me was that I used an open weight model and found 4 different zero days in 4 java libraries (they are all now fixed) and chained them to bypass validation of particular checks for digital signatures as my first finding (using LLMs) back in July. But, using open weights are either not allowed at every place or folks don't have hardware. We need powerful hardware to be able to defend. A lot of decisions have to made in various levels of our societies to make this happen. And I don't think we will not see much progress so easily.
On the other hand attackers using $25 and access to same capabilities to attack. Nation state level actors are now many more times efficient due to both brain power and money they can spend.
I hope the pace of finding new bugs will come down a bit in 2 years, but not sure how many of us will not be burnt out by then.
Note: many of you may ask me about technical details of the findings or what I am using specifically. That would require multiple other posts for me to explain, but you can just start poking with whatever you have. For example today morning's finding to another big project was using just Pi. No special sauce. Just poking and providing input to the model to guide it. But if I just go by numbers, then OpenAI models helped me the most in finding vulnerabiities and then to write exploits to demonstate to the maintainers. I am missing the capabilties to do proper pentesting though.