← all series

curl vulnerability disclosures

• Domain: vulnerabilities
• Role: discovery series
• Metric: vulnerabilities disclosed per quarter, split by whether the finder credit carries an AI marker
• Coverage: 2000–2026, partial through 2026-06-24
• Data: annual curl-by-year.csv (severity detail in the same file); quarterly curl-by-quarter.csv; per-finder curl-finders.csv
• Upstream: https://curl.se/docs/vuln.json (human-readable at https://curl.se/docs/security.html)
• Verdict: accelerating — 36 disclosures through 2026-06-24 annualize to roughly 75 against 9 in 2025 and a 13.1/year mean over 2014–2023

CVEs disclosed

The final bar is a partial quarter.

Definition

curl is a widely deployed C library and command-line tool. The project publishes a machine-readable record of every vulnerability it has ever disclosed, with a severity rating set by the maintainers and a credit string naming who reported it.

A "discovery" in this series is one disclosed vulnerability, counted in the quarter of its publication date. It is a disclosure count, not a count of bugs introduced or bugs remaining. An "AI-marked" disclosure is one whose credit string names an AI system or method, or names an AI-security employer; the two signals are recorded separately in curl-finders.csv and combined into the annual and quarterly tables' ai_attributed column.

Facts

• 2026 (through 2026-06-24): 36 disclosures; 15 AI-marked, 21 other
• 2026 annualized: roughly 75 disclosures
• prior rate: 9 disclosures in 2025; a 13.1/year mean over 2014–2023
• 2026 quarters: 2026-Q1: 10 · 2026-Q2: 26
• ai-band severity (2026): 12 of 15 AI-marked disclosures rated Low (80%), none High or Critical; 10 of 21 other disclosures rated Low (48%)
• severity drift: 2010–2022 disclosures were 18% Low and 28% High or Critical; 2023–2025 non-AI disclosures were 67% Low

The severity figure cuts the same rows by curl's own rating: one grid per finder-credit cohort, years across, ratings down, every cell printing the number of disclosures it holds. Shading is scaled within each panel, as the note on the figure states.

The collection-wide cumulative index redraws this series as cumulative disclosures to date:

Method

The CSVs are built by fetch.py, which reads curl's JSON record and buckets disclosures by publication year and quarter. The shared classifier in ../../lib/credits.py reads three independent signals off each FINDER string: whether it names an AI system or method (Big Sleep, Mythos, Claude, "agent"), whether it names an AI-security employer (Aisle Research, AntAISecurityLab, OpenAI, Anthropic), and whether it names fuzzing. The per-finder table curl-finders.csv records which band each credit falls in. The annual and quarterly tables keep a single combined ai_attributed column, true when a credit carries either AI signal; it describes the credit's text, not the method used.

figure.py calls the shared periodic_stacked() shape in ../../lib/families.py, drawing stacked quarterly bars from curl-by-quarter.csv: other_attributed in blue, ai_attributed in red, the final quarter outlined rather than filled and annotated with the annual table's data_through date. The axis is linear and January 2026 onward is shaded, as in every figure here. The severity figure comes from the same script through the shared severity_heatmap() shape, built from the annual table's severity columns and normalized within each panel. check.py recomputes the fact lines above from the CSVs.

Limitations

• credit text is not method evidence. Classification is by textual marker and errs in both directions: a researcher who used a model without saying so counts as human, and 14 of the 15 AI-marked credits of 2026 name only an employer.
• severity is compared across the combined AI band. The Low-severity shares above are computed against ai_attributed, so they describe reports from AI-security researchers rather than reports with a corroborated AI method.
• 2026 is a part-year, and curl publishes in batches at releases, so the quarterly bars show batch timing as well as a discovery rate.
• no denominator of effort. A credit records who reported, not how much search anybody spent.
• severity is one small team's judgment applied over twenty-six years, and may not be consistent across that span.
• triage load moved with the count. The Register reports that curl closed its bug bounty in January 2026 after AI-assisted submissions reached about 20% of entries while genuine yield fell [@stenberg2026bounty].

AI attribution

15 of the 36 disclosures of 2026 (through 2026-06-24) are AI-marked. One credit names a system:

"Andrew Nesbitt (powered by Mythos)" — curl credit string for one 2026 disclosure, vendored in curl-finders.csv, read 2026-08-14

The other 14 name an AI-security employer without stating a method: 9 credit Aisle Research ("Joshua Rogers (Aisle Research)" on 6, "Stanislav Fort (Aisle Research)" on 3), 3 credit AntAISecurityLab hackerone handles, 1 credits "Filipe Casal of Trail of Bits in collaboration with OpenAI", and 1 credits "Eunsoo Kim (Autonomous Code Security team at Microsoft)", all quoted from curl-finders.csv as read 2026-08-14.

2025 has 2 AI-marked disclosures: one system-naming credit,

"Google Big Sleep" — curl credit string for one 2025 disclosure, vendored in curl-finders.csv, read 2026-08-14

and one affiliation-only credit, "Stanislav Fort (Aisle Research)". No AI marker appears in any credit string before 2025, as of the 2026-06-24 snapshot. Stanislav Fort of Aisle Research appears in both curl and OpenSSL finder tables (../cyber-openssl/) [@googlebigsleep2024; @anthropicmythos2026; @aisle2026].

Sources

• [@bloomberg2026recordflaws] — press reporting connecting the 2026 disclosure records to AI systems; the counts here are built from curl's primary record, not from that reporting.
• [@googlebigsleep2024] — Google's vendor self-report on Big Sleep, the system named in the 2025 credit string.
• [@anthropicmythos2026] — Anthropic's Mythos preview, the system named in the 2026 "powered by Mythos" credit.
• [@aisle2026] — Aisle Research vendor commentary; Aisle researchers hold many of the AI-credited curl and OpenSSL findings.
• [@stenberg2026bounty] — The Register's report of the January 2026 bug bounty closure, cited for the limitation above.
• [@sherry2021fast] — measured heterogeneity of improvement rates across algorithm families; the published base rate for declining yields on a fixed target.
• Sibling series: OpenSSL and Firefox are the other fixed codebases with named finders; all software: disclosed counts CVE records across all software; OSS-Fuzz is the no-AI automation baseline over codebases that include curl.