al-ice.ai

curl 8.21.0 — AI-Powered AISLE Platform Finds 6 CVEs Including 25-Year-Old mTLS Flaw in Record Release

Security by al-ice.ai Editorial

AI relevance: AI-powered security platform AISLE used model-agnostic LLM analysis to discover 6 CVEs in curl/libcurl — more than any other research group — in the largest single curl security release ever, demonstrating that LLM-driven vulnerability research is now producing real-world findings in foundational infrastructure at scale.

What Happened

  • curl 8.21.0, released June 24, 2026, patches 18 CVEs — the most vulnerabilities ever fixed in a single curl version, with 276 bug fixes across 500+ commits from 100+ contributors.
  • The avalanche began May 11 when curl founder Daniel Stenberg announced that Anthropic's Mythos AI model had identified a single CVE in curl. That disclosure triggered an unprecedented flood of AI-assisted security reports.
  • AISLE, an AI-powered model-agnostic security platform, claimed 6 of the 18 CVEs — more than any other organization. The next-closest AI-powered group found 3. Researchers using Anthropic and OpenAI models found 1 each.
  • The crown jewel: CVE-2026-8932 — a 25-year-old mTLS connection reuse flaw first shipped in curl 7.7 (March 22, 2001). The bug allows authentication bypass when a client certificate changes but the connection is reused without re-authentication.
  • AISLE's other findings include credential confusion in netrc handling (CVE-2026-8926), a SASL double-free (CVE-2026-8925), use-after-free in socket callbacks (CVE-2026-9080), SSH host validation bypass (CVE-2026-9547), and HTTP/2 stream dependency use-after-free (CVE-2026-10536).
  • Beyond the 6 CVEs, AISLE also disclosed three additional memory safety issues via HackerOne: a heap out-of-bounds read in urlapi, and use-after-free/double-free bugs in HSTS handling.
  • Several vulnerabilities affect only libcurl — not the command-line tool — meaning they exist deep inside embedded products (containers, CI/CD pipelines, SDKs, automotive systems) where end users have no visibility and no direct patch path.

Key CVEs Found by AISLE

  • CVE-2026-8932 (Low) — mTLS connection reuse: authentication bypass after client cert change. Shipped since curl 7.7 (2001).
  • CVE-2026-8926 (Low) — netrc credential confusion: wrong user's password selected for the same host.
  • CVE-2026-8925 (Medium) — SASL double-free: GSASL context freed twice in protocol flows.
  • CVE-2026-9080 (Low) — Multi-socket use-after-free when curl_easy_pause() called inside socket callback.
  • CVE-2026-9547 (Low) — SSH host validation: rejected server key types accepted via libssh backend.
  • CVE-2026-10536 (Low) — HTTP/2 stream dependency tree use-after-free during handle cleanup.

Why It Matters

This is a concrete demonstration of AI-powered vulnerability research producing results in foundational infrastructure — not toy benchmarks. curl runs on 30+ billion devices. The fact that a 25-year-old authentication bypass survived until an AI model flagged it should reset expectations about what manual auditing alone can achieve in large, complex C codebases. However, several of these findings were rated Low severity, and the AI-finding signal-to-noise ratio remains an open question for broader application.

What To Do

  • Upgrade to curl/libcurl 8.21.0 immediately — especially environments relying on mTLS, SASL authentication, proxy configurations, or HTTP/2 and HTTP/3.
  • Audit embedded libcurl deployments: IoT devices, containers, CI/CD pipelines, and SDKs may not have a direct patch path.
  • Note upcoming removals: NTLM, SMB, TLS-SRP, and local crypto implementations are deprecated.
  • For AI security teams: AISLE's methodology demonstrates model-agnostic LLM analysis as a viable supplement to traditional fuzzing and manual audit of C codebases.

Sources