Last updated: September 5, 2026 · By Vishal Swami, Founder & Lead AI Reviewer, AISagely
Six curl CVEs after OpenAI and Anthropic came back with zero is the short version of what curl creator Daniel Stenberg disclosed in late August 2026: a security startup called AISLE found six real, fixable vulnerabilities in curl within a week of Anthropic's Mythos and OpenAI's Codex Security scanning the exact same codebase and turning up nothing. All six were confirmed, assigned CVE numbers, and patched in curl 8.22.0, according to AISLE’s own writeup of the disclosure.
Short answer: AISLE, a security startup, found six low-severity curl CVEs (CVE-2026-80229, -80230, -80231, -80255, -82208, -82209) in the last week of August 2026, days after OpenAI's Codex Security and Anthropic's Mythos scanned curl and reported zero confirmed vulnerabilities between them. All six were patched in curl 8.22.0, released September 2, 2026. It's a real data point on how differently AI security tools perform on the same target.
I write about AI coding and security tools for a living, and this story caught my attention because it's a rare case where several AI vendors pointed their tools at the exact same, well-documented codebase, in the same month, and the results are public enough to compare. In my testing, I also ran a small version of the underlying bug pattern through a plain review prompt myself, just to see how far a five-minute ask gets you compared to a dedicated scanner. Here's what actually happened, what it means if you're evaluating AI code-review tools, and where the hype in this space is outrunning the results.
What actually happened
curl is one of the most audited pieces of open-source software that exists — Stenberg has pointed out it's been through OSS-Fuzz, Coverity, CodeQL, and multiple paid human audits for years. In May 2026, Anthropic ran its Mythos security model against roughly 176,000 lines of curl's C source and reported five "confirmed" vulnerabilities. When Stenberg's team reviewed the findings, four turned out to be false positives or non-issues, leaving one genuine low-severity bug, which shipped as a fix in curl 8.21.0. Stenberg wrote afterward that the surrounding hype "appeared to be primarily marketing," with no evidence Mythos meaningfully outperformed prior AI tools run against the same code.
Then, on August 24, 2026, Stenberg said publicly that after running frontier AI systems from both major labs against curl, "Mythos says it can't find any more… Codex security shows an empty list." Zero confirmed findings from either OpenAI's or Anthropic's tooling, on a codebase that had already been through a public round of AI-assisted review months earlier.
How AISLE found what the frontier labs missed
The next day, August 25, 2026, Stanislav Fort of AISLE pointed his company's autonomous code-analysis system at the same curl repository. It produced 29 reports. Over the following days, curl's own security team reviewed them and confirmed six as genuine, CVE-worthy vulnerabilities — all rated Low severity, and all fixed in the curl 8.22.0 release on September 2, 2026:
- CVE-2026-80229 — OpenSSL provider use-after-free in pooled TLS connections
- CVE-2026-80230 — OpenSSL certificate pinning bypass
- CVE-2026-80231 — native CA store connection reuse issue
- CVE-2026-80255 — secure cookie attribute bypass involving a tab character
- CVE-2026-82208 — wolfSSL CA-cache hit overriding a verification callback
- CVE-2026-82209 — domain-scoped public-suffix list cookie handling bug
The headline bug, CVE-2026-80229, is a good example of the class of issue involved: libcurl can attach an OpenSSL provider context to an easy handle without taking an ownership reference on it, and if a pooled connection using that context outlives the handle, later reuse of the connection touches freed memory. It's the kind of lifetime bug that's easy to miss in a single-file read and much easier to catch if you're specifically tracing object ownership across a connection pool, which is closer to how a dedicated security-analysis agent works than how a general-purpose coding assistant works.
I tried a version of the same bug myself
To get a feel for the gap between a one-shot prompt and a purpose-built scanner, I wrote a stripped-down, 20-line version of the CVE-2026-80229 pattern — a struct that stores a raw pointer to a provider context, a cleanup function that frees it, and a pooled connection struct that can outlive the object that set it up — and ran it through Claude with a generic "review this C code for memory-safety issues" prompt, the kind of one-line ask most developers actually use day to day. It correctly flagged the dangling pointer and named the failure mode as a use-after-free tied to object lifetime across the pool, which is the right diagnosis for that isolated snippet. It also called out two unrelated lines as "worth double-checking" that weren't real problems in my toy example, which lines up with the false-positive pattern Stenberg described in Mythos's five-to-one drop after human review.
The honest caveat: that's a 20-line synthetic snippet with the bug in plain view, not 176,000 lines of production code with the bug buried three call sites deep in a connection pool. A quick prompt and a dedicated autonomous scanner are not the same product, and this story is really about the gap between them.
AI security scanners on curl, compared
| Anthropic Mythos | OpenAI Codex Security | AISLE | |
|---|---|---|---|
| When it ran | May 2026 | August 2026 | August 25, 2026 |
| Reports produced | 5 "confirmed" findings | 0 | 29 reports |
| Confirmed CVEs | 1 (low severity, curl 8.21.0) | 0 | 6 (all low severity, curl 8.22.0) |
| False-positive rate | 4 of 5 findings rejected | Not applicable, no findings | 23 of 29 not escalated to CVE |
| Approach | Frontier model reviewing source | Frontier model reviewing source | Purpose-built autonomous scanning agent |
| Public track record on this target | One prior round, modest results | First public test, zero results | New entrant, best public result to date |
The pattern worth noticing isn't that AISLE is flawless. Its 23 non-escalated reports out of 29 didn't clear curl's bar for a CVE either. It's that a system built specifically for vulnerability discovery, run by a team that does nothing else, outperformed general frontier-model security features from two of the best-funded AI labs in the world, on the same target, in the same month.
What you'll need to run this kind of review yourself
You don't need an enterprise security budget to try a lighter version of this. You'll want read access to the codebase you're reviewing, a model with a large enough context window to hold the relevant files rather than one function in isolation, and a human reviewer who actually understands the code. Every AI tool in this story needed one to separate real bugs from noise. If you're working in an agentic coding environment day to day, my guide to Claude Code usage tracking is a reasonable starting point for seeing what a model is actually doing across a review session, not just what it reports at the end.
Example prompts you can copy
- Ownership tracing: "Trace the lifetime of [struct/object] in this file. Does anything hold a reference to it that could outlive its cleanup function? Show me the specific call paths."
- Pool-reuse check: "This object is stored in a connection or resource pool for reuse. Walk through what happens to any pointers it holds if the object that created them is freed while it's still pooled."
- False-positive filter: "Before reporting this as a vulnerability, explain how an attacker would actually trigger it, and rate your confidence from 1-10 that this isn't a false positive."
- Second-pass request: "Re-review your own findings above. Which ones would you drop if you had to keep only the ones you're most confident are real?"
Common mistakes to avoid
The mistake I see most often is treating "an AI scanned our code and found nothing" as a clean bill of health. Codex Security and Mythos both came back empty on curl, and curl still had six real bugs in it. The absence of findings from one tool says more about that tool's blind spots on that codebase than it says about the code itself. Second, don't assume a higher report count means a better tool. AISLE produced 29 reports to get 6 real CVEs; Mythos produced 5 to get 1. Report volume without a human triage step is just more noise to sort through, not more security. Third, don't skip the human review step to save time. Every legitimate CVE in this story only became a CVE after curl's own security team checked the AI's work by hand. Fourth, don't compare tools based on a single run against a single codebase, including this one. curl is unusually well-audited already, so a tool that struggles here might still catch plenty in a codebase that's never had a real security review.
Tools that make this easier
If you want to see the flip side of this story, an AI code-review feature that shipped a real vulnerability instead of catching one, my writeup of GitHub Copilot Autofix and the Snowflake Jira breach is a useful contrast. For a case where AI-assisted bug hunting genuinely paid off at scale, Google fixed more Chrome bugs in June, thanks to AI covers a much larger, better-resourced version of the same idea. Z.ai’s security disclosure, where GLM-5.3 flagged 2,436 bugs, is worth reading next to this one for a sense of how wildly report volume swings between vendors. If you're weighing which AI coding assistant to trust with a security-adjacent task day to day, my comparisons of ChatGPT Codex and Cursor vs. Copilot cover the tools most developers are actually choosing between, and my deep dive on scanning Hugging Face datasets for secrets is a good next read if code security is the part of this story that interests you most.
My take
Six curl CVEs after OpenAI and Anthropic came back with zero is a good reminder that "AI security tool" isn't one product with one performance level. It's a category with a wide spread, and the vendor's name on the box tells you less than the track record does. I don't think this makes Codex Security or Mythos worthless; Mythos did find one real bug in May, and a frontier model with no dedicated security tuning finding anything in a codebase this heavily audited is a genuine result. But AISLE's showing is the one I'd point a team toward if vulnerability discovery in mature C code is the actual job, and I'd stay skeptical of any vendor's security claims until they've been checked against a target this well understood, by a maintainer this willing to publish the receipts.
Frequently Asked Questions
Is AISLE's tool free to try?
No pricing is listed publicly. AISLE offers a one-time "Snapshot" audit and an ongoing Enterprise product, and both require requesting access directly through AISLE rather than a self-serve signup.
How long did it take AISLE to find the six curl CVEs?
AISLE's system produced its 29 reports within a day of being pointed at curl on August 25, 2026. curl's security team then took about a week to review the reports and confirm six as genuine CVEs, which shipped in curl 8.22.0 on September 2, 2026.
Does this mean OpenAI and Anthropic's AI security tools are bad?
Not necessarily bad, but clearly beaten on this specific target. Mythos found one real bug in curl in May 2026, and Codex Security is a newer product with less public track record. The fair reading is that AISLE's purpose-built approach currently outperforms general frontier-model security scanning on a heavily audited codebase like curl's.
What is the easiest way to try AI-assisted code review myself?
Start with a plain-language ownership or lifetime-tracing prompt, like the ones above, against a file you already understand well enough to judge the answer. That gives you a sense of what a general model catches before you invest in a dedicated scanning product.
Were all six curl vulnerabilities serious?
No. All six CVEs were rated Low severity by curl's security team. They're real, fixable bugs, but none of them were rated as critical or high-impact issues.