The curl project has shipped version 8.22.0 containing fixes for six newly disclosed CVEs, all found by an autonomous AI security system built by the startup AISLE — and all discovered after frontier AI models from OpenAI and Anthropic had publicly reported that curl contained no further vulnerabilities. The comparison, documented in a blog post by AISLE researcher Stanislav Fort published Tuesday, is drawing attention across the security community as one of the clearest head-to-head tests yet of specialized AI security tools against general-purpose frontier models.
The baseline was unusually clean. On August 24, curl founder Daniel Stenberg publicly wrote that only three CVEs were pending for the next release, noting that Anthropic's Mythos model "says it can't find any more" and that OpenAI's Codex security system "shows an empty list" when asked to analyze curl, the ubiquitous data-transfer library deployed in more than 20 billion instances worldwide, from smart fridges to spacecraft. For more stories like this, see our AI security news coverage.
From zero to twenty-nine reports overnight
AISLE then pointed its own autonomous AI system at curl. The next day, Stenberg posted a one-line comparison that has since circulated widely: "Mythos: 0, Aisle: 29." Of those 29 reports, curl's security team reviewed six within days and judged them serious enough to merit public CVE designations in curl 8.22.0, which has now been released.
The six CVEs credited to AISLE and its researcher are:
- CVE-2026-80229 — OpenSSL provider use-after-free
- CVE-2026-80230 — OpenSSL pinning bypass
- CVE-2026-80231 — native CA store connection reuse
- CVE-2026-80255 — secure attribute bypass with tab
- CVE-2026-82208 — wolfSSL CA-cache hit overrides callback
- CVE-2026-82209 — domain-scoped public-suffix cookie
All six are rated Low severity. AISLE attributes that rating profile to curl's exceptional engineering maturity: after decades of auditing, the vulnerabilities that remain tend to hide in narrow configurations and subtle interactions between components, limiting practical impact. Three were reported on August 24, two on August 26, and one on August 27. By August 28, curl's pending CVE count had risen from three to ten, with six of the new findings coming from AISLE, according to the company's account.
Why this comparison matters
Security researchers have long dismissed much of AI-driven vulnerability discovery as theater — demonstrations staged against capture-the-flag challenges or benchmarks with known answers that may already lurk in model training data. This episode was different in one important respect: the analysis targeted current production code, and curl's own maintainers, not AISLE, decided both whether each finding was real and whether it warranted a CVE. Because Stenberg had publicly documented the frontier models' zero-result before AISLE ran its system, the comparison carried a timestamped baseline that is rare in vendor claims.
It is worth keeping the vendor context in view. AISLE sells AI code auditing as a product, and the blog post is explicit marketing for its "System over Model" thesis — the argument that specialized AI systems can beat general-purpose frontier models at real-world zero-day discovery. Commenters on Hacker News made the same observation, noting the post contains little technical detail about methodology, though discussion of the company's public code suggests an approach that combines LLM-guided targeting with traditional fuzzing: models identify interesting functions and generate test harnesses, then fuzzer findings are triaged. The company had previously disclosed six curl CVEs in June, including what it described as the oldest issue ever reported in the project.
The pattern may extend beyond curl
The most provocative signal came from outside the curl project. Greg Kroah-Hartman, the longtime maintainer of Linux's stable kernel releases, responded to Stenberg's post saying, "I'm seeing the same for Linux as well. No idea what Aisle is doing differently, but wow." If a specialized AI system is independently surfacing valid vulnerabilities in the Linux kernel that frontier models miss, the gap is not a quirk of one codebase.
For enterprise security teams, the episode lands amid a broader shift. Labs including OpenAI and Anthropic have invested heavily in positioning their frontier models for security work — OpenAI confirmed its Astra as a "critical" cybersecurity model just this week — and vulnerabilities are now regularly found by AI systems across open-source software. The curl result suggests the frontier models alone are not the whole toolchain: a general-purpose model reporting zero findings is not proof that findings do not exist.
The practical takeaways are already visible. Organizations running curl should update to 8.22.0, which contains fixes for all six CVEs. And for teams evaluating AI security tooling, the benchmark that matters is not how a model performs on canned challenges, but whether its findings survive review by the maintainers of real production code — the standard these six CVEs just met.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
