offlineaid-pack-builder is an early-stage project (0.x). Only the latest commit on main receives security fixes.

If you discover a security vulnerability, please report it privately via one of:

• GitHub Security Advisory (preferred — gives you a private channel with the maintainer)
• Email the maintainer (contact via GitHub profile)

Please do not open a public issue or pull request for security reports. The maintainer aims to acknowledge reports within 7 days and disclose responsibly after a fix is available.

In scope:

• packager.py — pack-builder CLI + archive contract
• agent.py — Gemma 4 compiler agent (PydanticAI + Deep Agents)
• extract.py — OCR extraction pipeline
• translate.py — translation pipeline
• scrape.py — Scrapy-based public-data extraction
• Makefile targets
• .oapack.zip archive format and SHA-256 sidecar contract

Out of scope:

• Third-party dependencies (report upstream — see pyproject.toml for the dependency tree)
• Hugging Face models pulled at pipeline time (report to model maintainers)
• On-device behavior in consumer apps that import .oapack.zip packs

• Zero runtime dependencies in the core packager.py — only the Python standard library (sqlite3, json, csv, struct, zipfile, hashlib, argparse). Optional extras (agent, ocr, scrape, upload) are isolated and only required for the pipeline stages that use them.
• No network at query time. Built .oapack.zip packs are self-contained SQLite databases consumed offline by the OfflineAid app.
• Archive contract is cryptographically validated. Every .oapack.zip carries a SHA-256 sidecar and is structurally + cryptographically verified at build time (_validate_archive_contract in packager.py) and at consume time (packager.py verify).
• Scrapy spiders are polite by default. ROBOTSTXT_OBEY=True, DOWNLOAD_DELAY=1.5s, CONCURRENT_REQUESTS_PER_DOMAIN=2. Only documented per-spider overrides exist, with rationale in code comments.
• Path traversal is validated. scrape.py and translate.py reject paths outside the working directory or system temp via _validate_path().

• Source-code provenance of the bundled examples/accc-scams/*.pdf — these are public documents from ACCC Scamwatch.
• The USER_AGENT string in scrape.py identifying this tool.
• Build-time logging to stdout (no PII expected; build logs are local).