Scientists developed a new AI model that automatically maps vulnerabilities to attack patterns using Large Language Models and identifies all relevant attack techniques—scientists from the Pacific Northwest National Laboratory report to Cyber Security News.
Imagine you’re the new manager of a large apartment building, and someone has stolen one of your keys—but you’re not sure which one. Was it to a first-floor apartment? The mail room? Maybe it’s a master key to all the units.
As you know, all locks are vulnerable, and you’ll need to change every lock to be completely secure.
But if you knew exactly which key went missing, you could target your efforts, changing just the relevant lock and eliminating the threat posthaste.
Multiply that problem thousands of times, and you’ll understand what cyber defenders grapple with.
There are more than 213,800 known “keys”—unofficial entry points into computer systems, better known as vulnerabilities or bugs—and they’re already in the hands of criminals.
There are likely many more that are not known. How can all the threats and attacks be tracked, prioritized, and prevented?
That’s impossible for any one person or team. While computer analysts share leads by feeding information into multiple databases, they don’t have a map of how adversaries might use most of those bugs to wreak havoc.
Now, a team of scientists at the Department of Energy’s Pacific Northwest National Laboratory, Purdue University, Carnegie Mellon University and Boise State University have turned to artificial intelligence to help solve the problem. The researchers have gathered three large databases of information about computer vulnerabilities, weaknesses, and likely attack patterns.
The AI-based model automatically links vulnerabilities to specific lines of attack that adversaries could use to compromise computer systems. The work should help defenders spot and prevent attacks more often and quickly. The work is open source, with a portion now available on GitHub. The team will release the rest of the code soon.
“Cyber defenders are inundated with information and lines of code. What they need is interpretation and support for prioritization.
Where are we vulnerable? What actions can we take?” Mahantesh Halappanavar, a chief computer scientist at PNNL who led the overall effort, said to Cyber Security News.
“If you are a cyber defender, you may deal with hundreds of vulnerabilities daily. You need to know how those could be exploited and what you need to do to mitigate those threats.
That’s the crucial missing piece,” added Halappanavar. “You want to know the implications of a bug, how that might be exploited, and how to stop that threat.”
The new AI model uses natural language processing and supervised learning to bridge information in three separate cybersecurity databases:
While all three databases have information crucial for cyber defenders, there have been few attempts to knit all three together so that a user can quickly detect and understand possible threats and their origins, and then weaken or prevent these threats and attacks.
“If we can classify the vulnerabilities into general categories, and we know exactly how an attack might proceed, we could neutralize threats much more efficiently,” said Halappanavar.
“The higher you go in classifying the bugs, the more threats you can stop with one action. An ideal goal is to prevent all possible exploitations.”
The work received the best paper award in November at the IEEE International Symposium on Technologies for Homeland Security. The work was funded by DOE’s Office of Science and PNNL’s Data-Model Convergence Initiative.
In addition to Halappanavar, the team includes first author Siddhartha Shankar Das of Purdue University, who was an intern at PNNL; former PNNL scientist Ashutosh Dutta, now at Amazon; Sumit Purohit of PNNL; Edoardo Serra of Boise State University and a joint appointee at PNNL; and Alex Pothen of Purdue.
In previous work, the team used AI to link two resources, vulnerabilities and weaknesses. That work, resulting in the model V2W-BERT, earned the team—Das, Pothen, Halappanavar, Serra, and Ehab Al-Shaer from Carnegie Mellon University—a best application paper award at the 2021 IEEE International Conference on Data Science and Advanced Analytics.
The new VWC-MAP model extends the project to a third category, attack actions.
“There are thousands upon thousands of bugs or vulnerabilities out there, and new ones are created and discovered every day,” said Das, a doctoral student at Purdue who has led the development of the work since his internship at PNNL in 2019. “And more are coming.
We need to develop ways to stay ahead of these vulnerabilities, not only the ones that are known but the ones that haven’t been discovered yet.”
The team’s model automatically links vulnerabilities to the appropriate weaknesses with up to 87 percent accuracy and links weaknesses to appropriate attack patterns with up to 80 percent accuracy.
Those numbers are much better than today’s tools provide, but the scientists caution that their new methods need to be tested more widely.
One hurdle is the dearth of labeled data for training. For example, very few vulnerabilities—less than 1%—are currently linked to specific attacks. That’s not a lot of data available for training.
The team fine-tuned pretrained natural language models using an auto-encoder (BERT) and a sequence-to-sequence model (T5) to overcome the lack of data and perform the work.
The first approach used a language model to associate CVEs to CWEs and then CWEs to CAPECs through a binary link prediction approach.
The second approach used sequence-to-sequence techniques to translate CWEs to CAPECs with intuitive prompts for ranking the associations. The approaches generated very similar results, which the cybersecurity expert on the team then validated.
“We’re putting this out there for others to test, to go through the vulnerabilities and make sure the model bins them appropriately,” said Halappanavar. “We really hope that cybersecurity experts can put this open-source platform to the test.”
In a resounding triumph for justice, U.S. District Judge Kathryn Kimball Mizelle has sentenced Vitalii…
Hackers are plotting to benefit from the generosity of Halloween, Thanksgiving, and Christmas shoppers using…
The LLMs (Large Language Models) are evolving rapidly with continuous advancements in their research and…
In the dynamic realm of mobile application security, cybercriminals employ ever more sophisticated forms of…
A recent campaign has been observed to be delivering DJvu ransomware through a loader that…
In a pivotal update to the Okta security incident divulged in October 2023, Okta Security…