Introducing Foundation V2

Improving on the best all around exploit prediction model

How do you improve on the best vulnerability scoring system in the business? Easy, you hire the smartest people in the biz, put them in a shared slack, and have them duke it out until they are shattering their own previous outstanding efforts. Today, you get to see the fruits of those conflicts in the form of Empirical Foundation V2, our prediction model for identifying global exploitation in the next 30 days.

How do we even define “Best”? Well here are some comparisons:

  • Nearly 300 more model inputs.

  • 3 new more exploitation sources.

  • 24% better prediction of exploitation in the next 30 days for all CVEs.

  • 10x (yes 1000%) better at identifying first time exploitation.

  • 2.2x better at identifying future exploitations that have been quiet for more than a month.

  • 1.9x better at recognizing when exploitation is likely to go quiet.

Those are some impressive numbers if I do say so myself, but we wouldn’t call ourselves “Empirical” Security if we didn’t actually have ways to back them up. So what does this mean for you, the yeoman vulnerability manager?

More Needles, less hay

This is a well worn analogy in our industry, since we have all come to the bitter realization that we can’t fix everything that hits the dashboard every day. Finding the high risk needles in the haystack of vulnerabilities is difficult, especially with the increasing number of published CVEs. Foundation V2 will change your exclamations from “Where the heck are all the needles?” to “What am I going to do with all these needles?” Observe:

The Haystack: Finding the Needles

The exploited CVEs are needles in a haystack of everything else. Each score hands you a pile to work through: how many needles are in it, and how big must it be?

Figure 1 With a fixed budget of 7,500 fixes (adjustable), 93% of the vulnerabilities you fix under Foundation V2 prioritization will be exploited in the wild within the next 30 days. Our previous model reached 60%. Meanwhile, CVSS reaches only 5% on the same budget.

Of course agrarian inspired visualizations of english idioms are good communication tools, we can also use the language of machine learning to really get the scope of exactly how well we are doing. Let’s look at the number of CVEs fixed vs the coverage of those CVEs being actual honest to goodness exploited in the wild in the next 30 days.

Coverage: share of the exploited CVEs you fixed

Foundation V2Foundation v1EPSS v5CVSSQualys QVSTenable VPRPerfect ordering: every exploited CVE first

Figure 2 That green line is about as close to perfect as we can get given the mess of that is vulnerability data. Everything that you fix is essentially exploited up to about 5k CVEs.

Score Fixes to
catch 50%
Fixes to
catch 80%
Fixes to
catch 90%
CVEs it
scores
Foundation V24,7457,8599,872100%
Foundation v14,74310,42929,486100%
EPSS v58,43541,79798,536100%
CVSS99,302211,452265,381100%
Qualys QVS40,610261,215319,20442%
Tenable VPR131,199278,795327,99424%

While we think Figure 2 does a pretty good job of conveying the performance in terms a data minded vulnerability manager would understand we can take a step further into the realm of ML measures where us Data Science sickos thrive using the industry standard PR curve.

Coverage and Efficiency

Foundation V2 (PR-AUC 0.957)Foundation v1 (PR-AUC 0.844)EPSS v5 (PR-AUC 0.530)CVSS (PR-AUC 0.043)Qualys QVS (PR-AUC 0.235)Tenable VPR (PR-AUC 0.168)

Figure 3 Coverage and efficiency. One of the standard way machine learning nerds measure model performance. Coverage is the percentage of exploited CVEs that are fixed. Efficiency is the percentage of fixed CVEs that are exploited.

Why show this esoteric curve? It’s because that’s what all the ML folks have decided is the best way to measure a classification model, specifically the area under those curves. It allows us to compute those hard numbers above. PR-AUC improvements in overall CVEs( 24% over Foundation v1), first exploitation detection (10x over Foundation v1), 2x on CVEs stopping exploitation, and 1.9x on old favorites returning.

Where the rubber meets the road

Data and curves are nice, but there are some real life examples that can demonstrate just how much better this model is doing for users. Here are three CVEs that saw exploitation shortly after publication that the Foundation V2 model scored correctly:

  • CVE-2026-94127 a F5 BIG-IP APM heap overflow was rated at 0.84 a day after publication in V2, whereas EPSS had it at a meer 0.01.

  • CVE-2026-87902 a WordPress core file inclusion vulnerability scored a 0.97 3 days after publication, whereas EPSSv5 lagged behind at 0.03

  • CVE-2026-76461 a Cisco Secure Email Gateway command injection vulnerability was scored immediately at 0.88 and has stayed high, compared to EPSSv5 which has capped out at 0.28

Just as important for a model like this are not just alerting on high risk vulns, but not panicking about vulns that pose little risk. Some older vulns that EPSSv5 insists are highly likely to be exploited foundation V2 correctly identifies as low likelihood exploitation (later born out by a lack thereof), these include:

  • CVE-2018-17456 a Git submodule command injection vulnerability

    • EPSSv5 scored at 0.97

    • Foundation V2 scored at 0.008

  • CVE-2020-11984 a Apache HTTP Server mod_proxy_uwsgi vulnerability

    • EPSSv5 scored at 0.90

    • Foundation V2 scored at 0.01

  • CVE-2019-3799 a Spring Cloud Config directory traversal vulnerability

    • EPSSv5 scored at 0.85

    • Foundation V2 scored at 0.005

Not just a score, a probability

Foundation V2, just like EPSSv5 and previous iterations of our Foundation model, is not just a number from 0 to 1, massaged and manipulated into a range that feels right, it’s a real gosh durn probability. That means you can reason about the Foundation V2 scores in any kind of risk model your heart desires and you can know that it’s not an approximate measure of risk, but a real component of any kind of cyber risk model. Concretely, of all the CVEs with a 5% chance of exploitation, ~5% of them will be exploited in the next 30 days. This holds true for all probabilities in our data. 

Calibration: Foundation V2

Perfect calibration 95% range Foundation V2

Figure 4 A well calibrated model. The horizontal axis here is the average score for CVEs in particular bin of scores. The vertical axis is the percentage of those CVEs in that bin that are actually exploited in a months time. If the model is outputting an accurate probability the dots should be on the diagonal (within the 95% confidence interval), which they are.

Stability, day to day, and when upgrading

A day in vulnerability management can be made or broken based on how scores for various findings shift on a day to day basis. One appeal of static scores like CVSS is that once a critical, always a critical, they aren’t going to sneak up on you. But a major cost with static scores (in addition to CVSS being a very bad measure of risk), is that they can’t be responsive to the changing risk landscape. Foundation V2 threads the needle¹ of being responsive while minimizing the jumps or declines in score that could have a vulnerability manager pulling their hair out. We did this through advancement and refinement of our input signals, particularly around exploitation history.

¹This one is not from a haystack. Different idiom.

In foundation v1 ~0.1% of CVEs (400!) experienced a day to day change of more than ten percentage points, with ~150 changing more than by more than 0.5. Big swings in scoring should happen sometimes (hot new exploit code, big malware campaigns), but perhaps foundation V1 was a little too sensitive to the tempest that is vulnerability data. Foundation V2 corrects this with more stable features and better identification of what really matters, with only an average of 150 CVEs changing their score by more than 10 percentage points and only 7 by more than 50 percentage points  day to day.

From one day to the next

Foundation V2 Foundation v1

Figure 5 We have to admit the old model tended to be a little sensitive to the chaotic landscape that is CVE exploitation. Through some fancy feature engineering, we’ve smoothed out the bumps as well as improved our performance. Most day to day changes are exceedingly small, and big changes only happen when they should.

It’s also notable that Foundation V2’s model performance improvements won’t do too much to disrupt the day to day operation of folks who are using Foundation V1. Most (80%) Foundation V2 scores are going to be within 1 percentage point of their foundation V1 value, and a very small percentage (0.5%) increase by more than 20 percentage points. Rest assured, switching over from Foundation V1 to Foundation V2 will not scramble your current prioritization list!

A solid foundation for assessing risk

Our foundation models are named “foundation” for a reason. In particular, they are our view of how to best assess exploitation across the plethora of sources we manage. But our wide, global view of the risk landscape is not perfectly tuned to individual organizations. That is why we offer Radiant models, which allow us to take this bedrock solid foundation and incorporate telemetry, controls, asset discovery, and more to build organizations an artisanal, personalized view of risk. Those models will be on even more solid footing with Foundation V2.

Next
Next

Visualizing the Vulnpocolypse