ahmedandaloes/security-decisions

Security decisions from NVD, KEV and EPSS

Typed, labelled decision cases for security triage, built for System 1 decision models (non-autoregressive: a state and typed questions in, calibrated probabilities out — see Laya and the Decision 3.0 family).

Each case is one vulnerability description plus up to ten typed questions, each with a label that an authority published — never an annotator's opinion:

question	type	label source
exploited_in_wild	noul	CISA KEV membership
exploitation_status	choice (none/poc/active)	CISA SSVC, via NVD
automatable	noul	CISA SSVC, via NVD
technical_impact	choice (partial/total)	CISA SSVC, via NVD
severity	score (4 levels)	CVSS base score, bucketed
attack_vector	choice	CVSS AV
privileges_required	choice (none/low/high)	CVSS PR
user_interaction	noul	CVSS UI
weakness_class	choice (10 buckets)	CWE, mapped
likely_exploited_30d	noul	EPSS >= 0.5

A question is labelled only when its source covers that CVE, so coverage differs per question and is reported in stats.json. Unlabelled questions are omitted rather than guessed.

What was removed, and why

Any description that states an answer out loud is dropped: "actively exploited", "in the wild", "proof of concept", "ransomware", and similar phrases. The label must not be readable off the text. Mere mentions of exploitation ("allows an attacker to exploit this") are kept, because they leak nothing about whether it was exploited.

Fields

• state — the NVD English description, verbatim
• questions — the typed question definitions to pass to the model
• expected — the label for each asked question (bool for noul, level index for score, option key for choice)
• _meta — CVE id, publication date, CVSS base score

Known limitations

• CISA KEV is a lower bound on exploitation: "not in KEV" means "not known to be exploited".
• The negative class is drawn from NVD publication windows, so date coverage is uneven; matching on year or on CVSS bucket when evaluating is strongly recommended, and that is how the model trained on it is measured.
• EPSS is a model's forecast, not a record of an event.
• CVSS and SSVC are judgements by CNAs and CISA; a model reproducing them is reproducing a rating, not discovering a truth.

Attribution

NVD (NIST), CISA KEV, FIRST EPSS, MITRE CWE. Built by the community; not affiliated with any of them.