Of the 45 security findings Anthropic reported to the curl project in early November, curl’s maintainers have confirmed seven as vulnerabilities, all of LOW severity, lead developer Daniel Stenberg wrote on Mastodon on 10 November. Another 37 were judged “just a bug” and one was “completely false”, he wrote.
By our arithmetic, that is 16% of the batch confirmed as vulnerabilities, while 44 of the 45 findings, or 98%, turned out to be real bugs of some kind. The reports are public as issues in curl’s GitHub tracker, where the project now handles LOW and MEDIUM vulnerabilities in the open.
Business Pill · Verification Cost
An evergreen Business Engineer explainer on why output that is cheap to make is not cheap to check. It teaches the general idea only and does not discuss this story.
The key insight: As we read it, the numbers separate two questions: whether an AI finding is a real bug, which 44 of the 45 were, and whether it is a security vulnerability, which seven were, all LOW, by the maintainers’ grading.
Stenberg first announced the batch on LinkedIn: “We received 45 new #curl vulnerability reports this morning (from Anthropic). I’m slowly moving them over to GitHub.” On 9 November he posted a link to “the 45 findings Anthropic reported to #curl last week”.
The 45 issues were opened between 5 and 9 November. Their titles include “Pinning bypass with TLS early data” and “HTTP/2 push on a DoH request: use-after-free in doh_probe_write_cb”, the tracker shows.
His tally followed a post that read: “for those who follow: yeah, we’re gradually raising the bar for what we accept as LOW severity #curl vulnerabilities.” Then came the count: “out of Anthropic’s latest 45 findings in #curl, we are at 7 confirmed LOW.”
Under the labels curl set out on 3 November, a finding marked as none is one the team deems to have no vulnerability risk, “or it is so low that it is considered to be less than severity LOW”; it “is treated as just a bug”. Confirmed vulnerabilities get a CVE and an advisory, published in sync with the next curl release, the post says.
One dot per finding: of the 45 security findings Anthropic reported to curl in early November, the maintainers confirmed 7 as LOW-severity vulnerabilities, judged 37 “just a bug” and found 1 completely false, as of 10 November (Daniel Stenberg on Mastodon; curl GitHub issues). Shares are our arithmetic.
What the AI Finds, in the Maintainer’s Words
In a reply on 10 November, Stenberg wrote that the tools “do find all kinds of bugs, they’re in particular good at finding similar things to what have been found in the past”. He added: “They tend to go deeper and deeper into the weeds the longer we play this game.”
In a reply on 9 November, he wrote: “The AIs are not good enough to produce fixes for these reports.” Separately, on 10 November he wrote that in the 27 days since the previous curl release the project had merged 218 bugfixes by 49 authors, with 19 CVEs in the queue for publication. His posts do not say how many of those CVEs come from this batch.
How It Compares With May
This is not curl’s first Anthropic-model scan. In a May blog post, Stenberg wrote that a Claude Mythos report, run for curl by someone with access to the model, listed five “Confirmed security vulnerabilities”. After review, his team was left with one confirmed vulnerability, three false positives and one finding deemed “just a bug”.
In that post he wrote that “the big hype around this model so far was primarily marketing”, while also saying that AI-powered code analyzers “are significantly better at finding security flaws and mistakes in source code than any traditional code analyzers did in the past”.
Other AI tools have a longer record on the same code. Stenberg wrote in October that engineers at Aisle had pointed their AI tooling at curl and that this had produced thirty-five published curl vulnerabilities and CVE IDs. As we reported in October, Cohere has agreed to buy Aisle.
How curl’s maintainers graded two scans: in May, a Claude Mythos report’s five findings it called confirmed became one confirmed vulnerability, one bug and three false positives (Daniel Stenberg’s blog, 11 May 2026); in November, Anthropic’s 45 findings became 7 confirmed LOW vulnerabilities, 37 bugs and 1 completely false (Stenberg on Mastodon, 10 November). Shares are our arithmetic.
The Ledger Gap
Security firm VulnCheck, which tracks Anthropic’s disclosures, wrote that Anthropic’s public ledger listed 37,966 findings and 217 fixes as of 2 November, and that the ledger claims a 94.2% true positive rate, up from 90.8% at launch. By our arithmetic, the fixes equal 0.57% of the findings.
VulnCheck counted 34 curl reports published at the time it wrote: 19 marked as not a vulnerability, six confirmed as low severity and nine awaiting assessment. It wrote that Anthropic’s OSS Scanner “shifts validation work to maintainers”.
Anthropic said in its 8 October post that it had found over 29,000 candidate vulnerabilities but had manually reviewed about 6,000, and that it remains “bottlenecked on our human capacity to validate these findings”. Its opt-in scanner’s reports are “fully model-generated, without human review or triage”, the company said, as set out in our piece on the free scans.
The two counts measure different things, by our reading. Counting every real bug, 44 of curl’s 45 findings held up, or 98%; counting only vulnerabilities, seven did, or 16%. The sources we read do not set out how Anthropic’s ledger defines a true positive. Anthropic’s own exploit-test figures are in our piece on its filing.
The Structural Read
As we read the posts, the cost of an AI security report lands on whoever must grade it: Stenberg wrote that the project merged 218 bugfixes by 49 authors in 27 days, and that the AIs are not good enough to produce fixes for these reports.
As we read VulnCheck’s figures, the gap it describes is between findings and fixes: 217 fixes against 37,966 ledger findings as of 2 November, 0.57% by our arithmetic, while Anthropic said its own human review covered about 6,000 of over 29,000 candidates.
Daniel Stenberg, curl lead developer, Mastodon, 10 November 2026
“out of Anthropic’s latest 45 findings in #curl, we are at 7 confirmed LOW.”
Three Implications
FOR OPEN-SOURCE MAINTAINERS curl grades every report in public, with labels that separate confirmed vulnerabilities from findings treated as just a bug, under the process Stenberg set out on 3 November.
FOR ANTHROPIC VulnCheck wrote that the OSS Scanner shifts validation work to maintainers; Anthropic said its scanner reports are fully model-generated, without human review or triage.
FOR SECURITY BUYERS Stenberg wrote in May that AI-powered code analyzers are significantly better at finding security flaws than traditional code analyzers, while calling the Mythos hype primarily marketing.
The analysis says of Anthropic: “It is helping define the rules, permissions, and boundaries around frontier intelligence.”
As we read it, the curl batch shows the other side of that layer: access to a frontier security model produces reports, and the open-source project receiving them sets its own rules for what counts as a vulnerability. The analysis predates this story and does not discuss it.
What Is Not Established
Stenberg’s posts we read do not name the Anthropic model or programme that produced the November batch. VulnCheck wrote that Anthropic rated one finding medium and the rest low; we did not see that rating in Stenberg’s own posts. The tally is his count as of 10 November, and assessments may change, as his labels allow.
CVE identifiers and advisories for the seven confirmed issues had not been published in the material we read. We did not contact Anthropic or the curl project for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
A Business Engineer analysis of the governance layer of the AI stack: who gets access to frontier intelligence, under what conditions, and for which purposes.
Anthropic reported 45 security findings to curl in early November. The project’s lead developer counted seven confirmed LOW-severity vulnerabilities, 37 plain bugs and one false report as of 10 November. By our arithmetic, 16% of the batch were vulnerabilities and 98% were real bugs, and the work of sorting one from the other fell to curl’s maintainers.
A note on sourcing. We read Daniel Stenberg’s Mastodon posts through the Mastodon API, his LinkedIn post, the curl GitHub issues tagged with Anthropic report identifiers through the GitHub API, his blog posts of 11 May, 14 October and 3 November 2026, Anthropic’s 8 October post and VulnCheck’s Glasswing update, all in full. VulnCheck’s ledger figures are its own reading of Anthropic’s ledger, attributed to it.
The percentages are ours. We first saw the tally in a relay on X and went to Stenberg’s own posts. We did not contact Anthropic or the curl project.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
