[SizzleUnrlsd]

Summary

Findings from four new repositories under tests-project/: Vulnerable_fastapi (six documented vulnerabilities in report/REPORT.md), InsecureGram (eleven in its README), DjangoGoat on its broken branch, and yt-fts, a real click command-line tool. Each fix is traced to a line they exposed.

• Tests first (test: define FastAPI, SQL position, deserialization, redirect and CLI behaviour), implementation follows.
• A method on what a project function returns resolves through its summary even when the call graph already derived a project symbol for it: get_conn().execute(...) in another file was app.database.get_conn.execute, unknown to the models, and hid the FastAPI login injection.
• Sink.positions: SQL sinks read the statement only, so a tainted value in the parameter tuple of a parameterised query is not an injection, while the credential kind keeps every argument. Without it the fix above would have flagged every parameterised query of Vulnerable_fastapi.
• An annotated parameter denotes its class in the call graph, so db: Session = Depends(get_db) gives db.execute its sink and sqlalchemy.text is a sink whose result carries no further SQL taint: InsecureGram's search injection.
• Two new taint kinds with sinks and detectors: DESERIALIZATION (pickle, marshal, shelve, dill, jsonpickle, unsafe yaml loaders; insecure-deserialization, critical) and REDIRECT (Flask, Werkzeug, Django, FastAPI and Starlette redirects; open-redirect, medium). Existing expectations built on yaml.load(input()) gain the deserialization finding.
• Command-line entry points: click.command, click.group, the commands of a group and of a Typer application make their parameters argv input; a module-level function decorated by a symbol is an instance of that symbol, so @cli.command() resolves. argv input carries every kind but PATH, a tool being expected to open the paths it is given; sys.argv follows.

On the four repositories: the FastAPI login injection and the guarded download traversal of Vulnerable_fastapi; the injection, deserialization, redirect, command execution, SSRF and upload of InsecureGram; two guarded redirects in DjangoGoat; nothing wrong in yt-fts. Still missing from their ground truth: Django views registered in urls.py and Model.objects.raw, next pass; template XSS, IDOR and CSRF, outside static analysis of Python.