Do not open a public issue. Use GitHub's private reporting flow: Security → Report a vulnerability on https://github.com/rlaope/jeval/security/advisories/new.

Please include:

• the affected version or commit,
• a minimal reproduction (a JSONL snippet plus the command is usually enough),
• the impact you believe it has,
• whether the issue can be triggered by untrusted input.

You can expect an acknowledgement within seven days. If a fix is warranted, it is released as a patch version with an advisory that credits you unless you prefer otherwise.

jeval is a local CLI. It makes no network calls of its own, has no telemetry, no accounts, and no server component. Everything it reads and writes stays on the machine that runs it:

• .jeval/records.jsonl holds decision records — treat it as production data, because it often contains user text, labels, and model identifiers.
• costs.yaml and thresholds.yaml hold business figures: what a wrong automation costs, what an escalation costs.
• report.html is a self-contained file with no external assets, but it does contain your data. It is ignored by .gitignore for that reason.

.gitignore excludes all of these paths by default. If you find a path through which user records, cost figures, or generated reports can be committed to a repository, that is a security report — it is exactly the class of leak this project must not have.

In scope: data leaks from the paths above, unsafe deserialization of ingest input, resource exhaustion from a crafted record file, incorrect handling of untrusted field values in the HTML report.

Out of scope: the accuracy, calibration, or honesty of any model that jeval measures; results produced from deliberately falsified input data; and anything requiring an attacker to already control the machine running jeval.