Here’s a demo video: https://www.youtube.com/watch?v=L6ZTu86qK8U#t=7
We are building Jazzberry to help you find bugs in your code base. Here’s how it works:
When a PR is made, Jazzberry clones the repo into a secure sandbox. The diff from the PR is provided to the AI agent in its context window. In order to interact with the rest of the code base, the AI agent has the ability to execute bash commands within the sandbox. The output from those commands is fed back into the agent. This means that the agent can do things like read/write files, search, install packages, run interpreters, execute code, and so on. It observes the outcomes and iteratively tests to pinpoint bugs, which are then reported back in the PR as a markdown table.
Jazzberry is focused on dynamically testing your code in a sandbox to confirm the presence of real bugs. We are not a general code review tool, our only aim is to provide concrete evidence of what's broken and how.
Here are some real examples of bugs that we have found so far.
“Authentication Bypass (Critical)” - When `AUTH_ENABLED` is `False`, the `get_user` dependency in `home/api/deps.py` always returns the first superuser, bypassing authentication and potentially leading to unauthorized access. Additionally, it defaults to superuser when the authenticated auth0 user is not present in the database.
“Insecure Header Handling (High)” - The server doesn't validate header names/values, allowing injection of malicious headers, potentially leading to security issues.
“API Key Leakage (High)” - Different error messages in browser console logs revealed whether API keys were valid, allowing attackers to brute force valid credentials by distinguishing between format errors and authorization errors.
Working on this, we've realized just how much the rise of LLM-generated code is amplifying the need for better automated testing solutions. Traditional code coverage metrics and manual code review are already becoming less effective when dealing with thousands of lines of LLM-generated code. We think this is going to get more so over time—the complexity of AI-authored systems will ultimately require even more sophisticated AI tooling for effective validation.
Our backgrounds: Mateo has a PhD in reinforcement learning and formal methods with over 20 publications and 350 citations. Marco holds an MSc in software testing, specializing in LLMs for automated test generation.
We are actively building and would love your honest feedback!
Our approach with Jazzberry is specifically focused on the dynamic execution aspect within the PR context. I am seeing that by actually running the code with the specific changes, we can get a clearer signal about functional errors. We're very aware of the need to demonstrate our ability to find those high-severity/exploitable bugs you mentioned, and that's a key metric for us as we continue to develop it.
Given your background, I'd be really interested to hear if you have any thoughts on what approaches you think might be most promising for moving beyond the false positive problem in AI-driven bug finding. Any insights from your work at MIT would be incredibly valuable.
We think that LLMs are able to capture semantic bugs that traditional software testing cannot find in a hands-off way, and ideas from both worlds will be needed for a truly holistic bug finder.
That seems like a waste of resources to perform a job that a static linter could do in nanoseconds. Paying to spin up a new VM for every test is going to incur a cost penalty that other competitors can skip entirely.
Our focus with the dynamic sandbox execution is aimed at finding bugs that are much harder for static analysis to detect. These are bugs like logical flaws in specific execution paths and unexpected interactions between code changes.
how many test cases do you have? how do you score the responses? how do you ensure random changes by the people who did almost all of the work (training models) doesn't wreck your product?