One AISLE curl Audit Produced 24 CVEs. 18 Ship Today

TL;DR: AISLE's 2026 curl audit has produced 24 CVEs across two releases. Six were fixed in curl 8.22.0, and eighteen more have just been fixed in curl 8.23.0. The audit, which kicked off after Codex Security, ZeroPath, and Mythos found no additional issues in curl, builds on a year-long relationship that has so far yielded 35 CVEs and 160 additional bug fixes.

Today, curl released version 8.23.0, fixing 24 previously unknown security vulnerabilities. AISLE discovered 18 of them, or 75% of the total. These findings, the result of an audit that delivered 24 CVEs across two separate releases, bring AISLE’s 12-month total of curl findings to 35 CVEs and 160 bugs.

We audited curl after its founder, Daniel Stenberg, reported that Anthropic's Mythos, OpenAI's Codex Security, and ZeroPath had found no additional vulnerabilities in curl. The following day, Daniel posted the result of the first batch of our findings plainly:

Six AISLE findings cleared curl's security review in time for the curl 8.22.0 release. We covered that first result in AISLE Discovered Six curl CVEs After OpenAI and Anthropic Found Zero. Daniel later asked us to treat the full body of work, spanning the initial run, follow-on analysis and two releases, as a dedicated AISLE audit of curl. This audit comes at the end of a year of collaboration in which AISLE provided both reproducible findings and useful patches for dozens of vulnerabilities.

What AISLE Found in curl 8.23.0

curl is one of the world's most mature, scrutinized and heavily tested software projects. Its maintainers estimate that it runs in roughly 30 billion installations. It has been developed in public for nearly three decades and examined by generations of security researchers, fuzzers, static analyzers and, more recently, frontier AI systems.

Many of the obvious bug classes in curl have been attacked for years. The vulnerabilities that remain increasingly hide in interactions between connection reuse, credentials, redirects, protocol state and rarely used configurations. This is true of AISLE's 18 findings in this release.

Credentials and identity crossing the wrong boundary

Several findings involved credentials or authenticated state surviving transitions where curl should have discarded or revalidated them:

• CVE-2026-86409: stale AWS SigV4 headers sent to an unintended destination
• CVE-2026-86410: redirect to SFTP or SCP bypasses host verification
• CVE-2026-86411: FTP connection reuse inherits a client certificate
• CVE-2026-86413: Negotiate credentials sent to a redirected origin
• CVE-2026-86414: host credentials leaked to an HTTP proxy
• CVE-2026-86415: automatic client-certificate connection reuse mix-up
• CVE-2026-94265: Authorization: Negotiate sent to a foreign origin
• CVE-2026-94266: password leak through CURLINFO_EFFECTIVE_URL

These are not all literally the same bug, but they share a theme: a piece of security-sensitive state remained attached to a handle or connection after the surrounding trust context changed.

Persistent state behaving unexpectedly

Four findings involved cookies or HSTS state being erased, truncated or loaded incorrectly:

• CVE-2026-86412: server-set cookie erases the cookie jar
• CVE-2026-87862: HSTS cache import drops a longer-lived child entry
• CVE-2026-87863: HSTS response erases persisted entries
• CVE-2026-92386: server-set cookie aborts cookie-jar loading

Protocol and command-line edge cases

Another group involved protocol-specific or command-line behavior:

• CVE-2026-87864: IMAP APPEND with CRLF conversion enables command injection
• CVE-2026-92388: FTP wildcard write callback mix-up
• CVE-2026-92389: TFTP direction confusion
• CVE-2026-92390: -J no-clobber guarantee bypassed through 103 Early Hints

Memory safety and availability

The final two findings affected process stability:

• CVE-2026-92387: server response omission causes a never-ending loop
• CVE-2026-92391: recursive curl_mime_free calls can exhaust the stack

Why These Bugs Survived

Most of these vulnerabilities do not sit inside one obviously dangerous line of code, which makes them difficult for many conventional security tools to detect. Instead, they emerge across time and state and require more complex reasoning to uncover. For instance:

• A connection enters a pool under one identity and is later considered for reuse under another.
• A redirect changes the destination but leaves authentication state behind.
• A cookie or HSTS entry is loaded, updated and persisted through several layers of logic.
• An unusual protocol response changes which callback receives data.

Each of the above functions can look reasonable on its own. The security failure appears only when the system mishandles context changes. While fuzzers excel at exercising parsers and static analyzers catch many local memory and data-flow errors, several CVEs in this release required tooling that can follow a security invariant across files, API calls, transfers, and protocol transitions.

AISLE's system searches for those broken invariants at codebase scale. It generates relevant context, traces state across components, proposes concrete failure sequences, and then subjects each candidate to skeptical triage.

How the Findings Were Validated

AISLE doesn’t simply report plausible bugs. Instead, the system attempts to disprove any potential finding before we send it to a maintainer.

Our AI checks candidate findings against current code, documentation, supported configurations, and the project's stated security model. Surviving findings are reproduced with a concrete test case or sequence of API calls. It then traces affected versions, assesses the security impact, and prepares a patch or regression test where possible.

Finally, an AISLE security researcher reviews the complete report before submission to make sure that we don't overwhelm the project maintainers with lower-quality findings.

The curl project then performs its own independent review. Its security team decides whether the behavior is real, whether it violates curl's security policy, how severe it is, and whether it merits a CVE. curl, not AISLE, makes the ultimate call on assignment and responsible disclosure.

This is the distinction between useful AI security research and slop. The point is not to produce more reports, but to help defenders fix more vulnerabilities before they are exploited.

A Year of Work with curl

The audit behind the last two releases is only one part of AISLE's relationship with curl. Since late 2025, AISLE has discovered dozens of curl CVEs and 160 additional security and performance bugs. Earlier findings included CVE-2026-8932, the oldest curl vulnerability reported to date. Introduced in curl 7.7 in 2001, it allowed libcurl to reuse a connection authenticated under the wrong client identity after mutual TLS settings changed.

The reason why this CVE survived for 25 years in a heavily analyzed and hardened codebase is instructive. The vulnerability was neither a parser crash nor an obvious missing bounds check. Rather, it only appeared across a sequence of connection-pool decisions and configuration changes. In other words, it was an error in the system’s reasoning about identity over time.

That is exactly the sort of pattern AISLE was built to find.

System Over Model

We call this approach "system over model", a phrase I coined to describe a simple idea: in applied AI, the system around the model can matter as much as the model itself. The curl audit was not a one-off research demo for us. Findings from work like this feed directly into AISLE's vulnerability discovery, triage, and remediation platform.

The underlying models will keep changing. Some large models perform extremely well on one security task and fail unexpectedly on another. That’s why AISLE's system selects and combines models, including our own custom fine-tuned models, builds codebase context, filters false positives, validates findings, and generates fixes. The system around the model is what turns raw capability into results that maintainers can use.

Update curl Now

AISLE recommends upgrading to curl 8.23.0 or applying the relevant patches for affected supported branches.

The same discovery engine behind all 195 of our findings in curl is available through AISLE Snapshot, a one-time security review that can run in the cloud, on-premises, or fully air-gapped.

Our thanks to Daniel Stenberg and the curl security team for their speed, rigor, and collaboration throughout the review.

See the complete AISLE curl record and AISLE's public CVE discoveries.