curl 8.23 ships 24 security fixes, most found with AI and verified by humans

curl 8.23.0 has shipped carrying 24 security fixes, and the story behind them is less about AI finding bugs than about what has to happen around the machines for their output to count. According to an analysis published on dev.to, most of the issues were surfaced with AI assistance — and every one of them passed through human reviewers before becoming a CVE.

curl 8.23.0 lands 24 security fixes

The release includes 24 security fixes, the dev.to piece reports. AISLE, a security research group, claims credit for 18 of them. The single issue rated High — a use-after-free in the HTTP/2 upload path — came from an Anthropic researcher using Anthropic's newly released open-source scanner. The curl team confirmed it within hours and shortened the release cycle from eight weeks to six to get the fix out.

Nine months from noise to signal

The contrast with January is the point of the story. Back then curl shut down its bug bounty after a flood of low-quality submissions; maintainer Daniel Stenberg said the goal was to stop rewarding poorly researched reports, whether AI-generated or not. In the program's final week, seven reports arrived and not one described an actual vulnerability.

Same project, same class of tools, nine months apart — one round produced enough noise to close a bounty, the other produced real CVEs. The dev.to author argues the difference was not the model but the process.

Where the human checks sit

According to dev.to, AISLE's workflow puts a security researcher on the complete report before anything is submitted. Curl's own security team then decides whether the behavior is real, whether it violates the project's security policy, how severe it is and whether it warrants a CVE. Two human gates, one on each side, before anything counts.

Stenberg has described AISLE's submissions as detailed, accurate work backed by skilled humans, and says the project is now fixing vulnerabilities at a higher rate than ever before. The January reports skipped the first gate entirely: someone ran a tool, got something plausible-sounding and forwarded it, shifting the cost of verification onto a small maintainer team.

Anthropic's raw firehose

On October 8, per the dev.to write-up, Anthropic opened an opt-in scanning service for open-source maintainers that explicitly ships raw output: fully model-generated, with no human review or triage. Anthropic says its own six months of scanning produced more than 29,000 candidate vulnerabilities, of which roughly 6,000 were manually reviewed.

The dev.to author does not consider this a bad thing — the service is free, opt-in and honestly labeled — but notes that it moves the triage burden onto whoever enrolls. The first human check becomes the maintainer's job.

Two failure modes recur, the piece argues. Agents invent findings: asked to look for a vulnerability, a model will produce something plausible, with a title and sometimes a severity attached. And agents declare areas safe, when a reasoned negative only means "stop spending time here," not that the audit is complete. Someone who knows the code well enough to catch both remains the bottleneck that did not get cheaper.

Why it matters

The curl episode is a working answer to the question every project with a public codebase now faces: what do you do with AI scanner output? curl's approach — reward nothing unverified, accept reviewed work eagerly — turned a tooling story into a security win, with 24 fixes shipped on a shortened cycle. The practical lessons generalise: a finding without a reproduction is a lead, not a finding; triage costs more hours than scanning; discarded results deserve a written reason, because trust assumptions change; and a clean scan of one code path says nothing about the rest. Running the tools is the cheap part now. The human who knows the code is the expensive part, and the part worth protecting.