parmar is an offline research harness for compression experiments. It is not a hardened archival format and should not be treated as one. Specifically:

• The archive format has no authentication and no encryption. The sha256 in the footer is an integrity check against accidental corruption, not a MAC — it is stored in the clear next to the data it describes, so anyone who can modify the payload can modify the hash to match.
• Decompression allocates memory driven by values read from the archive (token batches, and for zstd_22_long a window of up to 2 GiB). A hostile archive can make it allocate more than you want.
• parmar_core shells out to xz, zstd, gzip and bzip2. Argument vectors are built from a fixed internal table and never from archive contents, and every subprocess is launched as an argv list — never through a shell — so archive data cannot inject arguments. Tool paths come from PATH or from a fixed list of well-known install directories, so a writable directory earlier on PATH is the usual caveat.

Do not decompress parmar archives from untrusted sources, and do not use this format for anything where authenticity matters. If you need that, wrap it in something that provides it.

Please report security issues privately rather than opening a public issue:

• Use GitHub's Report a vulnerability (Security → Advisories), or
• email ronakparmar2428@gmail.com

Please include the version or commit, your platform and tool versions (python resources.py prints all of them), and a minimal reproduction. A crash or an unbounded allocation reachable from a crafted archive is in scope.

Expect an acknowledgement within about a week. This is a personal research project, not a funded product, so there is no formal SLA — but reports will be taken seriously and credited unless you prefer otherwise.

• The absence of authenticated encryption in the archive format (documented above, by design).
• Resource exhaustion from an archive you chose to decompress after being told not to trust it.
• Vulnerabilities in xz, zstd, gzip, bzip2, tiktoken, or numpy — please report those upstream. (If parmar invokes one of them in an unsafe way, that part is in scope.)