Stephen Shaffer, Laura Voicu
9 min
Organizations struggle to prioritize vulnerability remediation because they face thousands of new vulnerabilities annually while relying on static scoring systems that ignore organizational context. The authors propose a Bayesian framework designed to answer a specific question: How many exploit events should an organization expect given its unique exposure, and which remediation actions will most effectively reduce that risk?
To solve this, the researchers use exploit likelihood models (ELMs)—like the Exploit Prediction Scoring System (EPSS)—as a foundation. They then translate these global probabilities into local hazard rates. This involves three steps: first, quantifying the effectiveness of internal security controls using a combination of expert opinion (seeded as a prior) and operational data (telemetry or simulation) updated via Bayesian inference. Second, adjusting the vulnerability likelihood based on these control effectiveness measures and attack-vector alignment. Third, applying survival analysis to produce a daily hazard rate, which can account for the fact that exploitation risk typically decays as a vulnerability ages (modeled using a Weibull distribution).
The framework allows organizations to aggregate hazard rates from individual vulnerability instances up to specific hosts, business units, or the entire organization. By calculating the difference in aggregate hazard before and after a potential remediation action, the model produces a quantitative, defensible prioritization queue. Crucially, the authors demonstrate that prioritizing based on hazard reduction is not a proxy for volume; a high-risk vulnerability on an internet-facing host may rank significantly higher than hundreds of lower-risk vulnerabilities elsewhere, helping defenders allocate limited resources to the most impactful tasks.
This work moves vulnerability management away from qualitative, ordinal scoring (like traditional CVSS-based systems) toward a rigorous, quantitative framework rooted in survival analysis. By grounding the model in an organization’s own data and control efficacy, it provides a practical pathway for CISOs and security managers to translate abstract threat intelligence into measurable risk reduction. The approach is modular, allowing for future extensions into financial loss modeling and incident likelihood estimation, making it a foundation for data-driven cyber risk governance.
This paper presents a local exploit hazard model : a Bayesian framework that converts the global probabilities produced by an exploit likelihood model (ELM), such as the Exploit Prediction Scoring System (EPSS), into a daily exploit hazard rate for an organization's own assets. The model measures the exploit-prevention effectiveness of deployed controls as a probability distribution. That distribution is seeded from a subject-matter-expert opinion pool and updated through Beta-Binomial inference from telemetry, breach-and-attack simulation, or penetration testing, then applied to ELM scores by attack-vector alignment. The resulting per-vulnerability exploitation likelihoods are converted into hazard rates using standard survival-analysis techniques, supporting both a constant exponential hazard and a Weibull hazard whose shape parameter, calibrated from Known Exploited Vulnerabilities catalog timing, captures the empirical decay of exploitation risk as a vulnerability ages. Because hazards are additive under independence, per-vulnerability rates aggregate by summation up to host, network, business unit, and organization. Candidate remediation actions are simulated and ranked by projected hazard reduction, giving defenders a defensible, quantitative basis for prioritization under fixed capacity. Future work includes extensions for incident likelihood and financial loss modeling.
Sam: The authors acknowledge that coverage gap. Scripted attack libraries test known techniques, while penetration tests can explore novel attack chains but have narrower scope and run periodically. Expert judgment is meant to represent uncovered weaknesses. My caution is that this adds information about blind spots; it does not make biased logs unbiased.
Alex: Once they have that effectiveness estimate, how do they avoid crediting a defense against attacks it cannot actually stop?
Sam: They match defenses to the vulnerability’s attack vector, meaning the route through which exploitation occurs. A network-layer defense receives credit only against aligned network attacks. For each matching defense, the model retains the fraction of exploitation likelihood that the defense is estimated not to prevent.
Alex: Is that estimate really carried through as uncertainty, or does the final calculation quietly turn it into a single number?
Sam: The implemented calculation uses a single conservative summary of the effectiveness distribution. That keeps rankings cheap and deterministic, but discards uncertainty in the final hazard estimate. Multiple defenses are combined as if their prevention outcomes were independent. The authors also acknowledge that applying prevention once to an entire forecast window can understate remaining risk when repeated attempts matter.
Alex: So we now have a locally adjusted probability, under those assumptions. Why introduce another quantity instead of ranking that probability directly?
Sam: Because probabilities do not simply add across vulnerabilities or scale proportionally across time. The model converts them into hazard: an instantaneous exploitation rate, conditional on no exploitation yet. Under independence, hazards add. That gives the authors a common quantity they can aggregate across hosts, networks, or business units.
Alex: Does that rate stay fixed as a vulnerability gets older? A recently disclosed weakness and an old unpatched weakness might have different trajectories.
Sam: The paper offers both options. Its simpler model holds the rate constant. Its Weibull model, a survival model with an age-dependent rate, allows exploitation hazard to decline as vulnerabilities age. The authors calibrated that decline using roughly fourteen hundred usable timing records from the Known Exploited Vulnerabilities catalog maintained by the Cybersecurity and Infrastructure Security Agency.
Alex: Those records tell us when exploitation entered a catalog. Do they tell us when exploitation first happened?
Sam: Not necessarily. The authors use the interval between vulnerability publication and catalog addition, but addition may lag exploitation. The catalog is also curated around federal priorities, rather than randomly sampling exploitation. So the fitted decline is evidence from an imperfect timing proxy, not a direct measurement of every vulnerability’s exploitation history.
Alex: Let’s bring that back to a decision. Once a team has these rates, how does the framework tell it what to fix?
Sam: It simulates each candidate action by changing the relevant inputs and recalculating aggregate hazard. An action can be a patch, a defense improvement, or retiring a host. The model ranks actions by the hazard they remove. If effort estimates are available, it can instead rank reduction per unit of effort.
Alex: That makes the unit of comparison the work someone actually performs, rather than each vulnerability finding. Does the paper show a case where that distinction changes the ordering?
Sam: In its synthetic remediation example, an OpenSSL library upgrade removes more hazard than a macOS update, despite addressing fewer vulnerability instances on fewer hosts. The explanation is that the library supports internet-facing services, with weaker effective protection and higher exploitation forecasts. That illustrates the method’s logic; it is not a measured operational outcome.
Alex: And how much evidence supports the ranking itself? We have several uncertain inputs before we even reach the action list.
Sam: The authors tested sensitivity by taking a thousand random draws from their estimated effectiveness distributions. The experiment covered five defenses. With moderately informative estimates, the top five remediation actions retained their ranking. Both hazard models gave similar overall ranking agreement, while closely matched actions in the middle were less stable.
Alex: That supports stability under the uncertainty they represented. It does not establish that the predicted rates match what happens inside an organization. What is the biggest reason to keep that distinction visible?
Sam: This is a proposed localization framework, without reported prospective validation against an organization’s exploit counts. Independence is another consequential assumption: correlated vulnerabilities in the same software component can inflate the aggregate. Shared exposure can also cluster events. The paper explicitly leaves dependency modeling for future work, so stable rankings should not be mistaken for calibrated event forecasts.
Alex: Given that boundary, who should read the full paper, and where should they start?
Sam: Vulnerability-management researchers and security-risk modelers should read it as a framework to evaluate, not a validated incident predictor. Start with “Simulating Remediation Actions” to see the operational objective. Then read “Applying Control Effectiveness by Attack Vector” and “From Probability to Hazard” for the assumptions. Incident prediction and financial loss remain proposed extensions, not demonstrated outputs.
Alex: And what should someone outside that specialty carry back to their colleagues?
Sam: Compare actions by the risk they could remove, not just the vulnerabilities they close. This paper offers a way to make that comparison explicit. Whether its numbers deserve trust still depends on the evidence behind the local defenses and the assumptions connecting them to exploitation.
Alex: Thanks for listening.