AISLE Discovers 6 New CVEs in curl, Including the Oldest Issue Ever Reported
Author
AISLE Research TeamDate Published
.webp%3F2026-06-24T12%3A38%3A32.672Z&w=3840&q=100)
Curl runs on more than 30 billion devices. As one of the most important pieces of software in the world, it facilitates data transfers to and from operating systems, containers, appliances, CI pipelines, package managers, SDKs, cars, and is even on Mars inside NASA's Ingenuity Helicopter. Billions of users never run the curl command directly, but still depend on libcurl, the engine behind curl, through another product.
On May 11, 2026, curl founder and lead developer Daniel Stenberg announced that Anthropic’s Mythos model had found a single CVE in curl. His blog post unleashed a wave of research, which led to a flood of security reports to the curl project and, eventually, to the highest number of CVEs ever issued for a release of curl, 18.
AISLE led all security organizations with 6 of those 18 CVEs, plus additional valid findings, across curl and libcurl. The next-closest AI-powered organization received 3 CVEs, while researchers using Anthropic and OpenAI models found 1 each. This is further validation of our thesis: model-agnostic AI systems rival hyperpowered AI models, at a much lower price point, in any deployment environment.
All AISLE findings were responsibly disclosed to the curl project and were fixed in the June 24, 2026 release of curl 8.21.0. We urge everyone to update to the latest version.
Finding the Oldest curl Security Issue Ever Reported
Curl is of particular interest to security researchers: the easy bugs are long gone, and what remains is difficult to find: old protocol paths, state reuse, callback behavior, credential selection, and code paths that are easily forgotten about. That’s why we used AISLE’s autonomous vulnerability detection capability to find vulnerabilities in fall 2025, discovering 29 valid findings and 5 CVEs.
The 6 CVEs most recently identified by AISLE range from classic memory-lifetime issues to logic bugs in how libcurl decides whether a connection, credential, or host identity is still valid. They include CVE-2026-8932, the oldest curl vulnerability reported so far at over 25 years of age. Shipped in releases since curl version 7.7, it was first shipped on March 22, 2001.
A Summary of AISLE’s Findings
Notably, several issues only affect libcurl applications, not the curl command line tool. This means they affect the code embedded deep inside products where users do not know it is present, and where they become likely targets reachable through application behavior.
Finding | Area | What happened |
|---|---|---|
| curl could select a password belonging to a different user for the same host when the URL supplied a username but no password (credential confusion). | |
SASL authentication | curl could clean up and free the same GSASL context twice in SASL-enabled protocol flows (double free). | |
mTLS connection reuse | libcurl could reuse an existing connection even after client certificate or private key settings had changed (authentication bypass). | |
Multi socket callback lifecycle | calling | |
SSH host validation | with the libssh backend, SCP/SFTP transfers using a host-key callback could accept a server key type that should have been rejected (improper host validation). | |
HTTP/2 stream dependencies | resetting and then cleaning up a handle using HTTP/2 dependency options could cause libcurl to touch already-freed state (use after free). |
AISLE also reported several other curl bugs, including three memory safety issues:
- Use-after-free in `curl_easy_duphandle()` with HTTP/2 stream-dependency tree
- Heap-OOB read in urlapi `redirect_url()` via `CURLU_GUESS_SCHEME` + `CURLU_NO_GUESS_SCHEME` flow
- CURLOPT_HSTS_CTRL disables shared HSTS without share guard — use-after-free and double-free
Not every bug becomes a CVE, but these reports fall within the same category. They are all subtle edge cases in mature infrastructure code, especially around memory safety, state transitions, and esoteric API paths.
Bolstering the Case for Model-Agnostic Security Systems
The fact that AISLE claimed 6 of the 18 total findings in this release provides further support of our premise that well-engineered, model-agnostic systems rival high-powered frontier models on cybersecurity tasks.
Moreover, AISLE did more than simply discover vulnerabilities. Three CVEs were also patched using fixes generated by our platform. It goes to show that cybersecurity capability is jagged: for well-defined security tasks, smaller models can outperform much larger and more expensive LLMs. Notably, they can do so locally, completely on-premises, without making API calls.
The challenge is to match model capability and security needs. In other words, AI-native cybersecurity is not primarily a compute problem, but an engineering problem.
Engineering AI for Security with AISLE
AISLE’s end-to-end vulnerability management platform delivers autonomous security within your deployment constraints, from air-gapped networks to the cloud. If you want to see what AI will find in your codebase, talk to us.
Our sincere thanks to the curl project for their professionalism throughout the disclosure process. All our CVEs were reported and disclosed by Joshua Rogers of the AISLE Research Team.