empire-eval / scripts /ds_security.py
Lilambd's picture
Publish Lilambd Empire Eval
7f9a652 verified
Raw History Blame Contribute Delete
8.75 kB
from remote_util import latest, day_of, uniq
DS = 'Lilambd/security-daily'
LICENSE = 'CC BY 4.0'
STR = {}
KEYS = {}
SH = 'Security Daily'
def build(B):
root = B.store.root
f = latest(root, 'daily/cve_new/*.csv')
if f:
day = day_of(f)
KEYS[f] = ['cve_id']
c = B.store[f]
on = f'In {SH} (CVEs first published on {day})'
cu = uniq(c, 'cve_id')
for r in cu.loc[B.sample(cu.index, 40)].itertuples():
col = B.rng.choice(['cvss3', 'cvss3_severity', 'cwes', 'vendors', 'status'])
if getattr(r, col) != getattr(r, col):
continue
B.add('lookup', f'{on}, what is the {col.replace("_", " ")} of {r.cve_id}?', {'op': 'lookup', 'file': f, 'where': [['cve_id', '==', r.cve_id]], 'col': col}, tol={'abs': 0.05}, skills=['id_lookup', 'fresh_data', 'security'])
for sev in ['CRITICAL', 'HIGH', 'MEDIUM', 'LOW']:
B.add('count', f'{on}, how many new CVEs have CVSS v3 severity {sev}?', {'op': 'count', 'file': f, 'where': [['cvss3_severity', '==', sev]]}, skills=['filter', 'count', 'fresh_data', 'security'])
B.add('count', f'{on}, how many new CVEs are already in the CISA KEV catalogue?', {'op': 'count', 'file': f, 'where': [['kev', '==', 'True']]}, difficulty=2, skills=['filter', 'count', 'security'])
B.add('count', f'{on}, how many new CVEs have no CVSS v3 score yet?', {'op': 'count', 'file': f, 'where': [['cvss3', 'isna', None]]}, difficulty=2, skills=['filter', 'count', 'security'])
cw = c.cwes.dropna().value_counts()
for w in list(cw.index)[:8]:
B.add('count', f'{on}, how many new CVEs are classified exactly as {w}?', {'op': 'count', 'file': f, 'where': [['cwes', '==', w]]}, skills=['filter', 'count', 'security'])
B.add('rank', f'{on}, which CWE value appears most often among the new CVEs?', {'op': 'group_top', 'file': f, 'where': [], 'by': 'cwes'}, difficulty=2, skills=['group_by', 'security'])
B.add('rank', f'{on}, which vendor value appears most often among the new CVEs?', {'op': 'group_top', 'file': f, 'where': [], 'by': 'vendors'}, difficulty=2, skills=['group_by', 'security'])
f = latest(root, 'daily/patch_first/*.csv')
if f:
KEYS[f] = ['rank', 'cve_id']
p = B.store[f]
on = f'In {SH} patch-first list for {day_of(f)}'
for rk in range(1, min(len(p), 21)):
B.add('lookup', f'{on}, which CVE is ranked #{rk}?', {'op': 'lookup', 'file': f, 'where': [['rank', '==', rk]], 'col': 'cve_id'}, skills=['table_lookup', 'security'])
for r in p.head(30).pipe(lambda d: d.loc[B.sample(d.index, 12)]).itertuples():
B.add('lookup', f'{on}, why is {r.cve_id} listed (column why)?', {'op': 'lookup', 'file': f, 'where': [['cve_id', '==', r.cve_id]], 'col': 'why'}, skills=['table_lookup', 'security'])
B.add('lookup', f'{on}, what is the EPSS score of {r.cve_id}?', {'op': 'lookup', 'file': f, 'where': [['cve_id', '==', r.cve_id]], 'col': 'epss'}, tol={'rel': 0.001}, skills=['table_lookup', 'security'])
B.add('count', f'{on}, how many listed CVEs have known ransomware use?', {'op': 'count', 'file': f, 'where': [['kev_ransomware', '==', 'Known']]}, skills=['filter', 'count', 'security'])
B.add('rank', f'{on}, which listed CVE has the highest EPSS score?', {'op': 'argmax', 'file': f, 'col': 'epss', 'return': 'cve_id'}, difficulty=2, skills=['ranking', 'security'])
f = latest(root, 'daily/summary/*.csv')
if f:
KEYS[f] = ['ledger']
s = B.store[f]
for r in s.itertuples():
B.add('lookup', f'In {SH} summary for {day_of(f)}, how many rows in total does the ledger "{r.ledger}" hold?', {'op': 'lookup', 'file': f, 'where': [['ledger', '==', r.ledger]], 'col': 'rows_total'}, tol={'abs': 0}, skills=['table_lookup'])
f = 'snapshot/breaches.csv'
if latest(root, f):
KEYS[f] = ['id', 'org']
b = B.store[f]
bu = b[b.records.notna() & ~b.duplicated(['org'], keep=False)]
for r in bu.loc[B.sample(bu.index, 30)].itertuples():
col = B.rng.choice(['records', 'breach_date', 'disclosed'])
if getattr(r, col) != getattr(r, col):
continue
t = {'records': f'In {SH} breach snapshot, how many records were exposed in the {r.org} breach?', 'breach_date': f'In {SH} breach snapshot, on what date did the {r.org} breach happen (YYYY-MM-DD)?', 'disclosed': f'In {SH} breach snapshot, when was the {r.org} breach disclosed (YYYY-MM-DD)?'}[col]
B.add('lookup', t, {'op': 'lookup', 'file': f, 'where': [['org', '==', r.org]], 'col': col}, tol={'abs': 0}, skills=['table_lookup', 'security'])
for c in ['JP', 'US', 'FR', 'DE', 'GB']:
B.add('count', f'In {SH} breach snapshot, how many breaches are listed with country {c}?', {'op': 'count', 'file': f, 'where': [['country_iso2', '==', c]]}, skills=['filter', 'count', 'security'])
for src in b.source.dropna().unique():
B.add('rank', f'In {SH} breach snapshot, which organisation has the largest breach by records in the source "{src}"?', {'op': 'argmax', 'file': f, 'where': [['source', '==', src]], 'col': 'records', 'return': 'org'}, difficulty=2, skills=['ranking', 'security'])
for cls in ['Passwords', 'Email addresses', 'Phone numbers', 'Physical addresses', 'Government issued IDs']:
B.add('count', f'In {SH} breach snapshot, how many breaches list "{cls}" among the exposed data classes?', {'op': 'count', 'file': f, 'where': [['data_classes', 'contains', cls]]}, difficulty=2, skills=['filter', 'count', 'text_match', 'security'])
for org in ['Lilambd Crossing', 'Hibe Riverworks Ltd']:
B.add('unanswerable', f'In {SH} breach snapshot, how many records were exposed in the {org} breach? If there is no such breach, answer null.', {'op': 'absent', 'file': f, 'where': [['org', '==', org]]}, skills=['abstention', 'security'])
for yr in ('2025', '2026'):
f = f'snapshot/cve_signals/published_{yr}.csv'
if not latest(root, f):
continue
KEYS[f] = ['cve_id']
c = B.store[f]
cu = uniq(c, 'cve_id')
for r in cu.loc[B.sample(cu.index, 25)].itertuples():
col = B.rng.choice(['cvss3', 'epss', 'products', 'kev'])
if getattr(r, col) != getattr(r, col):
continue
B.add('lookup', f'In {SH} CVE signals, what is the {col} value of {r.cve_id}?', {'op': 'lookup', 'file': f, 'where': [['cve_id', '==', r.cve_id]], 'col': col}, tol={'rel': 0.001}, skills=['id_lookup', 'security'])
B.add('count', f'In {SH} CVE signals, how many CVEs published in {yr} are in the CISA KEV catalogue?', {'op': 'count', 'file': f, 'where': [['kev', '==', 'True']]}, difficulty=2, skills=['filter', 'count', 'security'])
B.add('count', f'In {SH} CVE signals, how many CVEs published in {yr} are flagged with known ransomware use?', {'op': 'count', 'file': f, 'where': [['kev_ransomware', '==', 'Known']]}, difficulty=2, skills=['filter', 'count', 'security'])
B.add('count', f'In {SH} CVE signals, how many CVEs published in {yr} have an EPSS score of at least 0.5?', {'op': 'count', 'file': f, 'where': [['epss', '>=', 0.5]]}, difficulty=2, skills=['filter', 'count', 'security'])
vend = c.vendors.dropna().value_counts()
for v in list(vend.index)[:10]:
B.add('count', f'In {SH} CVE signals, how many CVEs published in {yr} have vendor value exactly "{v}"?', {'op': 'count', 'file': f, 'where': [['vendors', '==', v]]}, skills=['filter', 'count', 'security'])
B.add('rank', f'In {SH} CVE signals, among {yr} CVEs with vendor "{v}", which has the highest EPSS score?', {'op': 'argmax', 'file': f, 'where': [['vendors', '==', v]], 'col': 'epss', 'return': 'cve_id'}, difficulty=3, skills=['ranking', 'security'])
for _ in range(4):
cid = f'CVE-{yr}-{B.rng.randint(990000, 999999)}'
B.add('unanswerable', f'In {SH} CVE signals, what is the CVSS v3 score of {cid}? If the CVE is not listed, answer null.', {'op': 'absent', 'file': f, 'where': [['cve_id', '==', cid]]}, skills=['abstention', 'security'])
f = latest(root, 'weekly/vendor/*.csv')
if f:
KEYS[f] = ['vendor']
w = B.store[f]
for r in w.itertuples():
B.add('lookup', f'In {SH} weekly vendor table for the week ending {day_of(f)}, how many new CVEs did {r.vendor} have?', {'op': 'lookup', 'file': f, 'where': [['vendor', '==', r.vendor]], 'col': 'cves_new'}, tol={'abs': 0}, skills=['table_lookup', 'security'])