AISLE says its researchers found 18 of the 24 CVEs fixed in curl 8.23.0
curl 8.23.0 shipped 24 security fixes nine months after the project closed its bug bounty over low-quality reports. The explanation on offer, a human check before each submission, is one auditor's reading.
The Engineer · Build desk
AISLE says its researchers found 75 percent of the release's fixes. Anthropic's own six months of scanning produced over 29,000 candidate vulnerabilities, and it manually reviewed at most about 21 percent of them.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Fixes in the release AISLE says its researchers found | 75% | 15 | |
| Candidate vulnerabilities from Anthropic's scanning, over | 29,000 candidates | in its own six months of scanning | 11 |
| Candidates Anthropic manually reviewed, at most about | 21% | of its own scan candidates | 16 |
What happened
- In the last week of curl's bug bounty, seven reports came in and none of them described a vulnerability.
- The release's single High-severity fix, for a use-after-free bug in how curl handles HTTP/2 uploads, came from an Anthropic researcher working with the company's new scanner for open source projects.
- curl's maintainers confirmed the bug in a matter of hours. To get the fix out, they shortened the release cycle to six weeks instead of eight.
- AISLE says a researcher reviews each complete report before submission, and curl's security team then rules on whether it is real, breaks policy, how severe it is and whether it gets a CVE.
Why it matters
- cost A maintainer who enrolls in a scanner with unreviewed output inherits the triage. The author says reading results and discarding most of them is where the hours go.
- constraint The pipeline moves only as fast as a maintainer who knows the code can confirm or reject. curl's hours-long confirmation depends on that, and a project without such a person has no equivalent.
- decision Projects deciding whether to accept AI-assisted submissions can set the intake bar at a reproduction. The author's rule is that a finding without one is a lead.
- precedent curl kept taking AI-assisted reports and filtered on verification. Other projects facing the same inbox can copy that filter without banning the tools.
The causal claim is the post author's. He audits smart contracts and wrote: "I don't think the difference was the model." [19] The January closure targeted a specific input. Stenberg said the aim was to remove the incentive to send "crap and non-well researched reports to us. AI generated or not." [2] By the author's account, curl kept accepting AI-assisted reports but stopped rewarding unverified ones [21].
What the post documents is a pipeline with two checks, one on the reporter's side and one on the maintainer's [14]. Stenberg said AISLE's work was "detailed and accurate findings backed up by a team of skilled humans." [9] The author says the January reports skipped the first check, so the cost of verifying moved from the reporter to a small maintainer team [17].
The High-severity bug is the weak point in that story. Its reporter used Anthropic's scanner. Anthropic opened that opt-in service for open source maintainers on October 8 and says its output "will be fully model-generated, without human review or triage." [10] The post does not say what checking the Anthropic researcher did before reporting. It credits the fast confirmation to curl's team knowing its own code [13]. We think the two-check reading holds for the AISLE reports and is unproven for this one.
The review gap shows in Anthropic's own numbers. Six months of its scanning produced over 29,000 candidate vulnerabilities, and it manually reviewed about 6,000 [11]. That is at most about 21 percent [16]. The author saw the same pattern on a small scale. In his last sweep on a competition codebase he ran six agents in parallel. Five returned reasoned negatives. The sixth produced passing proofs of concept that only rediscovered a finding his team already had [20].
For curl's result to transfer, the reporter side of a project needs someone who can reproduce a bug before sending it. The maintainer side needs to know the code well enough to rule in hours. Stenberg's rule asks for the first: never report a bug unless you actually understand it and can reproduce it [18].
AISLE's 18 of 24 is 75 percent of the release [15]. Without a count of how many reports AISLE submitted in all, the rejection rate behind that share is unknown. Stenberg now says the reviewed work is helping the project fix vulnerabilities "at a higher rate than ever before." [12]
What to watch
- Whether curl or AISLE publishes counts of reports submitted versus accepted, which would put the 18-of-24 share against a rejection rate.
- Whether maintainers who enroll in Anthropic's opt-in scanner report how much of its unreviewed output survives their own triage.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption
- Insufficient
- Hype gap+20
- Incentives50
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
curl shut down its bug bounty in January.
- [2]
Daniel Stenberg said the goal of closing the bounty was to remove the incentive for people to send "crap and non-well researched reports to us. AI generated or not."
- [3]
The last week of curl's bug bounty program brought seven reports and none of them described a vulnerability.
- [4]
curl 8.23.0 shipped with 24 security fixes.
- [5]
AISLE says its researchers found 18 of the 24 fixes.
- [6]
The only fix rated High, a use-after-free in the HTTP/2 upload path, was reported by an Anthropic researcher using Anthropic's new open source scanner.
- [7]
The curl team confirmed the High-severity bug within hours and cut the release cycle from eight weeks to six to ship the fix.
- [8]
AISLE says a security researcher reviews the complete report before submission; curl's own security team then decides whether the behavior is real, whether it breaks curl's security policy, how severe it is and whether it gets a CVE.
- [9]
Stenberg described AISLE's work as "detailed and accurate findings backed up by a team of skilled humans."
- [10]
On October 8 Anthropic opened an opt-in scanning service for open source maintainers and says its output "will be fully model-generated, without human review or triage."
- [11]
In its own six months of scanning, Anthropic got over 29,000 candidate vulnerabilities and manually reviewed about 6,000.
- [12]
Stenberg says the reviewed work is helping curl fix vulnerabilities "at a higher rate than ever before."
- [13]
The author says curl's team could confirm the High-severity bug in hours because they know their code.
- [14]
The author counts two human checks before anything became a CVE: one on the side that ran the tools and one on the side that owns the code.
- [15]
AISLE's 18 of the 24 fixes is 75 percent of the release.
- [16]
Anthropic manually reviewed at most about 21 percent of the candidates from its own scanning.
- [17]
The post's author says the January reports skipped the first human check, and the cost of verifying moved from the reporter to a small maintainer team.
- [18]
Stenberg's rule was to never report a bug unless you actually understand it and can reproduce it.
- [19]
The post's author wrote: "I don't think the difference was the model."
- [20]
In his last sweep on a competition codebase the author ran six agents in parallel; five came back with reasoned negatives and the sixth produced passing PoCs that only rediscovered a finding his team already had.
- [21]
curl did not stop taking AI-assisted reports; it stopped rewarding unverified ones.
- [22]
The author's advice: a finding without a reproduction is a lead, not a finding.
- [23]
The author says running the scanning tool is the cheap part now, and reading 200 results and throwing most of them away is where the hours go.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- dev.toAI helped find most of curl's 24 new CVEs. What made them real was the human check.
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Open Source SecurityFollow
- AI Vulnerability ScanningFollow
- Bug Bounty ProgramsFollow