CVE-2026-61343: Librebooking and GLM

A small open source project vs a cheap LLM…

Librebooking is a php based web application that provides:

flexible, mobile-friendly, and extensible interface for organizations to manage resource reservations

I knew of this application only because of CCDC https://www.nationalccdc.org/, where we do much training on thing such as securing php web applications.

This was one of my first attempts in a long list of web applications I wanted to explore using AI. I had a ChatGPT subscription at the time, but it was getting particularly aggressive with telling me that my activity was cyber related and that it was time for me to move on. I ended up using opencode https://github.com/anomalyco/opencode with a https://chat.z.ai/ GLM subscription. I also used some free DeepSeek V4 Flash as well.

This was also my first CVE, so this was an exciting process to go through.

PoC Code: https://github.com/nmagill123/CVE-2026-61343-poc-librebooking-rce/

The admin RCE was quite simple, it abuses an email template editor/upload functionality to add a web-shell to a spot in the web root that can execute php code.

The CVE description is as follows:

LibreBooking’s email template editor save action passes the submitted template name directly into the destination file path, allowing a remote attacker with administrator credentials to write an arbitrary file outside the template directory and execute code. Fixed in 5.1.0.

The code maintainers were fairly quick to get back to me and patch the web application, which is a good sign of open source project health. Check out the GitHub PoC for more information.

GLM?

Tired of getting “This chat was flagged for possible cybersecurity risk” in you terminal? Time to use AI models such as GLM or DeepSeek.

GLM is a Chinese based open weights AI model. It offers a flagship LLM experience at a fraction what the frontier models cost. Most of the industry is worried about the impact that models such as Fable or Mythos will bring to vulnerability research. This is undoubtedly true, but I think the “hype” around these models being problematic to security is not realistic as: they are far too expensive.

With a good personal skill set, something like DeepSeek or GLM can be very effective at finding vulnerabilities, especially because depending on your provider the inference speed can be very fast. This helps out with sub-agents, or other forms of AI automation.

My setup was quite simple: I told it to use docker, look for vulnerabilities in the code using OWASP Top 10 as a guide, spawn one sub-agent for a broad vulnerability category (i.e path traversal), and validate against a live stock setup.

Especially with cheaper models, they tend to hallucinate quite a bit, so do not get too excited when the model tells you it found an unauthenticated RCE.

Another thing I found is the models will find real vulnerabilities, but for obscure non-default setups. While some of these may be realistic enough to qualify as a vulnerability, some of these config changes that it would require to count as a vulnerability is far fetched.

I am sure that all of the Chinese AI providers are training models/surveilling your data, but I am particularly bullish on what they will do to the AI industry. I’m glad to see them cut popular AI companies model’s costs down for a small reduction in benchmarks. It was frustrating to me that I could not accomplish at home the scale that I can accomplish at work with a hefty AI budget when this all first started, but the gap is shrinking in affordability.

Also going back to my first statement, the safeguards are way less restrictive on these Chinese models. Good and bad, mostly good.

I think overtime this should allow for this type of intelligence to be more easily available to the masses.

Vulnerability Research right now…

Yeah, its not looking great. I have reached out to several OSS vendors who have told me they are not interested in triaging my security vulnerabilities in things that might’ve made headlines years ago. I do not particularly blame them, if you go to GHSA tabs on projects, depending on the repository settings, you will see thousands of “AI Slop” vulnerabilities.

I’m not sure that I will continue trying to find vulnerabilities using AI just because it is so time consuming to try to reach out, explain, validate findings, and compete with somebody else’s Sonnet/Opus in the process.

I do think that within a few years the landscape will improve and enough project owners will be per-emtivevly using AI to scan their own repositories on a regular cadence. That combined with efforts of well-meaning VR folks, this should help the overall scene. Companies will also undoubtedly using AI to do the same on closed source software.

Thanks for reading!