
Published in the evening, in a ticket by morning: Our security agents close a critical Next.js vulnerability
By 11 November 2026, around 95,500 security vulnerabilities with a CVE number had been published, about 300 a day on average. That is already twice as many as in all of 2025. No team reads all of that. That is why we built our own security agents: they continuously check our projects for vulnerabilities and start dedicated subagents for each task. Why we rely on agents for this is explained in our post on cyber defence with agents (in German). Here we show with a real case what this looks like: a critical vulnerability in Next.js on our own website.
More vulnerabilities, found faster
In 2025 there were around 48,000 vulnerabilities, 20 percent more than the year before. For 2026, the security organisation FIRST forecast around 59,000 in February and raised its forecast to around 66,000 in June. The actual number is already almost 50 percent higher.
AI is a major driver. In early 2026, Anthropic used Claude Opus 4.6 to find and validate more than 500 high-severity vulnerabilities in open-source software. Google's Big Sleep agent discovered a bug in SQLite that the project's established fuzzing tests had not found. But AI is not the only driver: according to FIRST, sources such as GitHub now report many times more than before. And the same tools help attackers turn a published vulnerability into a working attack faster.
For website and app operators, this means: anyone who goes through security advisories by hand once a month is too slow.
The case: remote code execution in Next.js
Vercel published two critical vulnerabilities in Next.js, both with a severity score above 9 (out of a maximum of 10). The first allowed an attacker to run their own code on the server without logging in, through the image optimisation with SVG files. Without a security update, our website ahoikapptn.com would have been affected too.
How it went
Late afternoon: Scan of our website. The scan agent checks every package, including indirect dependencies, against the GitHub Advisory Database, the OSV database, the US agency CISA's list of actively exploited vulnerabilities, and the vendors' security pages. Result: eight critical vulnerabilities, including the two in Next.js.
Next morning: The scan creates the tickets in Linear, listing each vulnerability, the affected package and the update that closes it.
Shortly after: A fix agent raises the affected dependencies to a secure version, closing the vulnerabilities. While testing, it lacks access to our private packages. Instead of carrying on untested, it opens the pull request as a draft and writes in the ticket exactly what a human needs to do.
Late morning: We set up the access. The agent tests in throwaway containers: build, home page, CMS login, editing a document, uploading an image. Everything works. Our AI review finds only one minor comment, which the fix agent resolves itself.
Midday: We approve the update, and a few minutes later it is live.
Found in the evening, the fix was ready as a pull request the next morning, without anyone on our team having to deal with it until then. The website was live with the fix less than 24 hours after the scan. Developer time: about 15 minutes for review and final testing, plus setting up the access once.
How our security agents work
The scan runs twice a day, in the morning and in the late afternoon. Whatever is published in the evening is in our tickets the next morning, and shortly after, a fix agent is already working on it. Up to the tested pull request, no developer is normally needed. A human only steps in for the approval.
The security agents coordinate the work and start dedicated subagents for each step:
- Inventory agent: records every project once a week with all packages, runtimes and servers.
- Scan agent: checks all sources for new vulnerabilities on every run. A fix is created for a CVSS score of 9 or higher, or when a vulnerability is actively exploited. If no update exists yet, the ticket says so, and the agent checks again on every scan.
- Fix agents: one per affected project. They update, test and open the pull request. Anything they cannot check themselves, such as the Node.js version on Vercel, they add to the ticket as a check for a human. If a vulnerability is actively exploited, they draft an email to the affected client.
On top of that, our AI review checks every pull request. And a human always makes the final decision: no agent publishes anything on its own.
Our security agents are currently in beta, and we keep extending them.
Key takeaways
- Know all your packages. Most vulnerabilities sit in libraries nobody installed on purpose.
- Prioritise by risk. Whether a vulnerability is actively exploited matters more than the CVSS score alone.
- Test before updating. A patch that takes the website down helps nobody.
- Automate searching and preparing.
Sources
- CRACI: CVE Publication Trends (as of 12 November 2026)
- FIRST: Vulnerability Forecast 2026
- FIRST: The 2026 Vulnerability Forecast Update
- Anthropic: Evaluating and mitigating the growing risk of LLM-discovered 0-days
- Google Project Zero: From Naptime to Big Sleep
- Vercel: GHSA-2464-fhr4-w9xw, Next.js RCE in image optimisation
- Vercel: GHSA-673q-mjm6-r5xr, Next.js RCE in "use cache"
- Ahoi Kapptn!: Cyber-Verteidigung im Zeitalter der Agenten (in German)