BUG TRENDS
Is the supply of bugs finite? · Confirmed vulnerabilities in mature C projects vs. AI bug-finding · v0.1.1 · updated 2026-11-20 06:17 UTC
247vulns 2026 YTD
(46 in 2025)
(46 in 2025)
40%AI-credited, last 12 mo
(nginx/curl/openssl)
(nginx/curl/openssl)
5.0 ymedian bug age 2026
(1.5 y in 2025)
(1.5 y in 2025)
RISINGannualised pace
vs last year
vs last year
The question
Does a codebase contain a finite number of bugs that AI can eventually exhaust, or is the supply effectively unlimited? We don't assume an answer. Even a frozen codebase may not have a fixed count, because what counts as a bug shifts as attackers, environments and tools change. Three competing hypotheses, and the patterns each would leave in the data:
| If bugs are effectively finite | If effectively unlimited | Staircase (bursts per capability jump) |
|---|---|---|
| Old-core (>5 y) findings spike then decline; median bug age falls | Findings flat or rising indefinitely, including in old code | Bursts aligned with capability milestones (red dashed lines), decay between |
Confirmed vulnerabilities per quarter — all six projects
Per project
Bug age at disclosure — the depletion test
Yearly summary
| Year | All vulns | High/crit | Credited set* | AI-credited | Median age (y) | Old core | Newer | Old share |
|---|---|---|---|---|---|---|---|---|
| 2016 | 109 | 45 | 66 | 0 | 4.4 | 23 | 40 | 37% |
| 2017 | 81 | 49 | 22 | 0 | 2.8 | 6 | 15 | 29% |
| 2018 | 66 | 21 | 24 | 0 | 3.6 | 9 | 13 | 41% |
| 2019 | 50 | 25 | 20 | 0 | 4.3 | 3 | 15 | 17% |
| 2020 | 33 | 16 | 14 | 0 | 5.9 | 7 | 2 | 78% |
| 2021 | 94 | 42 | 23 | 0 | 6.3 | 14 | 8 | 64% |
| 2022 | 77 | 19 | 37 | 0 | 2.9 | 15 | 20 | 43% |
| 2023 | 51 | 7 | 39 | 0 | 8.0 | 21 | 16 | 57% |
| 2024 | 68 | 28 | 27 | 0 | 2.4 | 9 | 18 | 33% |
| 2025 | 46 | 11 | 17 | 5 | 1.5 | 6 | 11 | 35% |
| 2026* | 247 | 41 | 177 | 70 | 5.0 | 87 | 89 | 49% |
* Credited set = nginx + curl + openssl, where finder and first-vulnerable version are published. 2026* = year to date.
Capability milestones
- 2016-12-01 — OSS-Fuzz launched (fuzzing baseline)
- 2022-11-30 — ChatGPT released
- 2024-06-20 — Google Project Zero Naptime (LLM vuln research)
- 2024-11-01 — Big Sleep finds SQLite bug
- 2025-08-08 — DARPA AIxCC finals
- 2025-10-01 — Aisle / AI-first research groups begin mass reporting to curl & OpenSSL
Latest 40
| Date | Project | CVE | Sev | Age y | Summary | Found by |
|---|---|---|---|---|---|---|
| 2026-11-17 | openssl | CVE-2026-107493 | low | 11.8 | Use-After-Free in PEM_X509_INFO_read_bio_ex() when appending to a caller-supplied stack fa | Filipe Casal (Trail of Bits) |
| 2026-11-17 | openssl | CVE-2026-124709 | low | 1.6 | Unbounded Memory Retention in QUIC Server for Rejected Streams | Bhabani Sankar Das |
| 2026-11-17 | openssl | CVE-2026-35190 | low | 1.6 | Out-of-Bounds Read in TLS 1.3 Key-Share Processing with Suite B Certificate Flag | Laurent Gaffie (secorizon.com) |
| 2026-11-11 | nginx | CVE-2026-123634 | medium | 19.0 | Use-after-free during response processing in resolver | |
| 2026-11-11 | nginx | CVE-2026-123007 | high | 11.1 | Memory corruption in HTTP/2 flow control | |
| 2026-11-11 | nginx | CVE-2026-121810 | medium | 3.5 | Use-after-free in HTTP/3 | |
| 2026-11-11 | nginx | CVE-2026-123633 | medium | 5.7 | Use-after-free in HTTP/2 | |
| 2026-11-11 | nginx | CVE-2026-118899 | high | 9.5 | Infinite loop while reading client request body over TLS 1.3 | |
| 2026-11-11 | nginx | CVE-2026-113425 | high | 19.0 | Use-after-free in mail proxy module | |
| 2026-11-11 | nginx | CVE-2026-120993 | low | 0.3 | NULL pointer dereference when using PROXY protocol version 2 | |
| 2026-11-11 | nginx | CVE-2026-125283 | low | 15.6 | Buffer overflow when using alias with try_files | |
| 2026-11-11 | nginx | CVE-2026-120967 | medium | 2.6 | SNI bypass when using client certificate validation in stream | |
| 2026-11-11 | nginx | CVE-2026-123640 | low | 8.3 | Insufficient validation in stream ssl_preread module | |
| 2026-11-10 | openssl | CVE-2026-88823 | low | 1.6 | AES-GCM-SIV AAD Forgery After Zero-length Incremental AAD Update | Jérémy Jean (ANSSI), Bhargava Shastry and Jonathan Bateman ( |
| 2026-11-10 | openssl | CVE-2026-107496 | low | 11.8 | Infinite Loop in CMS_encrypt() With an EC Group Order of 1 | Yinxi Liu |
| 2026-11-10 | openssl | CVE-2026-88822 | low | 1.6 | QUIC Uses Uninitialized Stack Bytes During Automated Peer Detection | Jaehun Lee |
| 2026-11-10 | openssl | CVE-2026-95691 | low | 1.6 | Retransmitted QUIC CRYPTO Frames Can Cause DoS or One-byte Disclosure | Filipe Casal (Trail of Bits) |
| 2026-11-10 | openssl | CVE-2026-95692 | low | 1.6 | Out-of-Bounds Read and Incorrect IV Length Handling in AES-GCM-SIV | Billy Brumley (Rochester Institute of Technology) |
| 2026-10-30 | sqlite | CVE-2026-119371 | medium | SQLite before 3.54.0 contains a stack exhaustion vulnerability that allows attackers to by | ||
| 2026-10-30 | sqlite | CVE-2026-119372 | medium | SQLite before 3.54.0 built with FTS3/FTS4 and SQLITE_ENABLE_FTS3_PARENTHESIS contains a us | ||
| 2026-10-30 | sqlite | CVE-2026-119373 | high | SQLite before 3.54.0 built with FTS5 contains a heap buffer overflow in fts5MergeRowidList | ||
| 2026-10-30 | sqlite | CVE-2026-119374 | high | SQLite before 3.54.0 built with FTS5 contains a heap buffer overflow in fts5DoSecureDelete | ||
| 2026-10-30 | sqlite | CVE-2026-119375 | medium | SQLite before 3.54.0 contains an information disclosure vulnerability that allows attacker | ||
| 2026-10-30 | sqlite | CVE-2026-119376 | medium | SQLite before 3.54.0 contains a stack exhaustion vulnerability that allows attackers to by | ||
| 2026-10-30 | sqlite | CVE-2026-119377 | low | SQLite before 3.54.0 contains an authorization bypass vulnerability in lookupName() that i | ||
| 2026-10-30 | sqlite | CVE-2026-119378 | medium | SQLite before 3.54.0 contains a use-after-free vulnerability in the R-Tree extension that | ||
| 2026-10-30 | sqlite | CVE-2026-119379 | high | SQLite before 3.54.0 built with the session extension contains a heap buffer overflow vuln | ||
| 2026-10-30 | sqlite | CVE-2026-119586 | high | SQLite command-line shell through 3.54.1 contains a command injection vulnerability in the | ||
| 2026-10-30 | sqlite | CVE-2026-119591 | medium | SQLite before 3.54.0 contains a stack exhaustion vulnerability in countOfViewOptimization( | ||
| 2026-10-14 | curl | CVE-2026-107929 | low | 2.2 | local interface connection reuse mixup | Maksim Hayder |
| 2026-10-14 | curl | CVE-2026-107928 | low | 4.0 | info leak uploading over WebSockets with GnuTLS | Luna with Claude; Anthropic |
| 2026-10-14 | curl | CVE-2026-94266 | low | 8.0 | password leak in CURLINFO_EFFECTIVE_URL | Joshua Rogers (Aisle Research) |
| 2026-10-14 | curl | CVE-2026-94265 | low | 0.3 | Authorization: Negotiate sent to foreign origin | Joshua Rogers (Aisle Research) |
| 2026-10-14 | curl | CVE-2026-92392 | high | 1.7 | HTTP/2 upload Use-After-Free | Luna with Claude; Anthropic |
| 2026-10-14 | curl | CVE-2026-92391 | low | 9.0 | curl_mime_free stack overflow by recursion | Joshua Rogers (Aisle Research) |
| 2026-10-14 | curl | CVE-2026-92390 | low | 16.7 | -J no-clobber guarantee bypassed with 103 early hints | Joshua Rogers (Aisle Research) |
| 2026-10-14 | curl | CVE-2026-92389 | low | 21.0 | TFTP direction confusion | Joshua Rogers (Aisle Research) |
| 2026-10-14 | curl | CVE-2026-92388 | low | 16.3 | FTP wildcard write callback mixup | Joshua Rogers (Aisle Research) |
| 2026-10-14 | curl | CVE-2026-92387 | low | 2.1 | never-ending loop via server response omission | Joshua Rogers (Aisle Research) |
| 2026-10-14 | curl | CVE-2026-92386 | low | 2.5 | server-set cookie aborts cookie jar load | Joshua Rogers (Aisle Research) |
Method & caveats
- Sources: NVD API (all projects), curl.se/docs/vuln.json (credits, introduced version), openssl-library.org vulnerability list (credits, affected branches), nginx.org security advisories (vulnerable range). Project pages override NVD, which lags months behind — NVD alone showed zero nginx CVEs for 2026.
- Bug age = disclosure date minus release date of the first vulnerable version. For OpenSSL the oldest listed branch is the oldest still supported, so ages there are a lower bound.
- AI credit is a pattern match on finder names (Aisle, Anthropic/Claude, Big Sleep, XBOW, AIxCC teams, AntAISecurityLab, …). It misses humans who used AI quietly and is approximate.
- CVE counts also reflect policy: projects differ in what they assign CVEs to, and curl/OpenSSL changed triage under the 2025–26 report flood.
- Rebuilt daily from source. Raw data: bugs.json.