EPSS Scores and AI-Discovered Vulnerabilities: Why Exploit Prediction Models Are Changing in 2026

EPSS (Exploit Prediction Scoring System) is maintained by FIRST.org and is currently in its third major version. The model produces a probability score between 0 and 1 representing the likelihood that a given CVE will be exploited in the wild within the next 30 days. The model is updated daily. EPSS v3 uses a gradient boosting model trained on historical CVE exploitation data, with features drawn from NVD metadata (CVSS score, attack vector, attack complexity, privileges required), threat intelligence signals (whether the CVE appears in threat actor toolkits or exploit kits), and proof-of-concept exploit availability in public repositories and databases. A score of 0.95 means the model predicts a 95% probability of exploitation within 30 days. A score of 0.02 means 2%. The median EPSS score across all CVEs is below 0.01, meaning most published CVEs are never actively exploited.

CVSS and EPSS answer different questions and are both necessary for effective triage. CVSS v3.1 measures the theoretical worst-case impact of a vulnerability: exploitability metrics capture how an attacker would reach and exploit the vulnerability; impact metrics capture what the attacker gains. A CVSS 9.8 score means the vulnerability is remotely exploitable, requires no privileges, and grants full system compromise. EPSS tells you the probability that exploitation will happen in the real world within 30 days based on current threat intelligence. The practical implication: a CVSS 6.5 vulnerability with EPSS 0.85 is operationally more urgent than a CVSS 9.0 vulnerability with EPSS 0.01. One is almost certainly going to be exploited; the other is theoretically catastrophic but unlikely to be targeted in the near term.

EPSS v3 improved on previous versions by incorporating a broader feature set and a more sophisticated model architecture. Key inputs include: public exploit availability in repositories including GitHub, ExploitDB, and Metasploit modules (the single strongest predictor of near-term exploitation); CVSS attack vector and complexity (remote, no-auth vulnerabilities have higher base rates); product popularity and installed base (CVEs in widely deployed software have higher exploitation rates); threat intelligence signals including mentions in dark web forums, threat actor reports, and security vendor advisories; and the age of the CVE (newly published CVEs see a spike in exploitation probability that decays over time). The model output is recalibrated daily as new exploitation data, threat intelligence, and exploit availability data flows in.

EPSS performs best for CVEs that have been in the ecosystem long enough to accumulate exploitation history and threat intelligence signal. For CVEs with published Metasploit modules, EPSS accurately identifies high exploitation probability. For CVEs in widely deployed software with large installed bases and active threat actor interest, EPSS correlates well with observed exploitation. For CVEs that are more than 90 days old, EPSS has had time to incorporate exploitation observations, public exploit publication events, and threat intelligence signals that sharpen its probability estimate. In these cases, EPSS provides a reliable complement to CVSS for prioritizing remediation queues.

AI-discovered vulnerabilities from Project Glasswing expose a structural limitation in EPSS. The model's strongest predictor of exploitation is public exploit availability, typically a Metasploit module or a GitHub proof-of-concept. For traditional CVD timelines, this signal appears days to weeks after the initial vulnerability is discovered by a human researcher. For Glasswing CVEs, Claude Mythos develops a working exploit during the same discovery session, before the CVE is even assigned. The exploit exists before the CVE has any EPSS history. When the CVE is published and EPSS scores it for the first time, it has no historical exploitation data and no public exploit in the repositories EPSS monitors, so it receives a low score. But a functional exploit exists in Anthropic's coordinated disclosure documentation. EPSS has no mechanism to incorporate this signal.

CVE-2026-4747 is a FreeBSD NFS kernel memory disclosure vulnerability discovered by Claude Mythos and disclosed through Project Glasswing's CVD program. CVSS score: 8.4 (high). At initial publication, EPSS will assign a low-to-moderate score because there is no historical exploitation data and no public Metasploit module. But Mythos developed a working exploit during its discovery session. The effective exploitation risk is higher than EPSS reflects because functional exploit code exists and the vulnerability affects a widely deployed operating system component. CVE-2026-5194 is a wolfSSL certificate chain forgery vulnerability with CVSS 9.1. Same pattern: the EPSS score at publication will not reflect the Mythos-generated exploit capability. Security teams applying raw EPSS to these CVEs will underweight them relative to their actual risk.

Security teams need to apply an explicit adjustment to their prioritization model for AI-discovered CVEs. The adjustment has three components. First, source tracking: when a CVE originates from an AI vulnerability discovery program (Glasswing, XBOW, or similar), apply a manual exploit-available flag regardless of whether a public Metasploit module exists. Second, disclosure-day treatment: treat AI-disclosed CVEs as having immediate elevated exploitation risk from day one rather than applying the standard exploitation ramp-up model that EPSS uses. Third, CVSS floor: for AI-discovered CVEs with CVSS of 7.0 or higher, apply a 30-day patch SLA regardless of EPSS score, because the low EPSS score reflects data absence rather than low risk.

The most effective triage approach combines four factors into a single priority score. Factor one: CVSS base score, weighted for attack vector (remote over local), attack complexity (low over high), and impact (confidentiality, integrity, availability). Factor two: EPSS probability, capped at a confidence floor for AI-discovered CVEs where EPSS underestimates risk. Factor three: explicit exploit availability flag covering public Metasploit modules, GitHub proof-of-concept exploits, and AI-generated exploit code documented in CVD disclosures. Factor four: asset exposure, specifically whether the vulnerable component is internet-facing or accessible from an adversary-reachable network segment. Combining these four factors produces a triage ranking that handles both traditional CVEs where EPSS is reliable and AI-discovered CVEs where it is not.

The Mythos Brief contains the complete adjusted prioritization model for AI-era vulnerability management, including the four-factor scoring formula, the EPSS integration guide for common vulnerability management platforms, and the AI-discovered CVE treatment policy template.