A security report you can verify three months later
I got tired of security reports that were just a claim. A scanner would hand me a JSON file full of findings, and six weeks later nobody could say for sure whether that was the file from the review, or which exact versi
I got tired of security reports that were just a claim.
A scanner would hand me a JSON file full of findings, and six weeks later nobody could say for sure whether that was the file from the review, or which exact version of the code it described. The findings were usually fine. The evidence wasn't.
So I rebuilt the scanner around the report instead of the other way round. It's called injection-audit. It does static taint analysis on Python and Bash - sources like sys.argv and os.environ, sinks like os.system and subprocess with shell=True - and then treats the output as an artifact rather than a log line.
Every report has an ID derived from the tree and the date, so the same repo on the same day gets the same ID. Each finding points to a file and line, quotes the offending source, and records a SHA-256 of the exact file that was read. A coverage section says how many files were looked at and which ones were skipped and why - a binary or a README doesn't just vanish from the count. Then the whole payload is signed with HMAC-SHA256, and the report records the key fingerprint but never the key.
That last part started as a small feature and turned out to be the point of the whole thing. A signed report means someone who wasn't in the room can check it later. No account, no dashboard, no service to trust. You need the report and the verify command.
I ran it on a deliberately broken six-file app to see what it would actually do. Eight findings: five in Python, three in Bash. One shell=True call with an f-string came back tagged as both CWE-78 and CWE-88, which is right - it's a command injection sink and the argument boundary is broken at the same time. A bash -c in the deploy script got flagged too. Four files were analysable; one binary and one README were skipped and counted as such.
The output makes no certification claim. It says so in a fixed disclaimer inside the report itself. These are findings and dated evidence, with the limits written down: taint tracking is intra-procedural, alias tracking through containers is limited, no JavaScript, no cross-module dataflow, no CI gating. I'd rather state that plainly than have someone discover it in the middle of an audit.
The engine is standard-library Python. No external calls, nothing to install beyond the package, no phone-home.
Code and a sample report are at github.com/derbaroser-eng/injection-audit. If you have ever had to defend a finding months after the fact, I would like to know what your report format looked like.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.