This is a research preview (v0.1.0). Only the current release is supported.

ABI is a research prototype for academic and experimental use. It is not intended for production deployment.

Known limitations relevant to security:

• wikitext_cache.py downloads data from Hugging Face datasets — use only in trusted network environments
• Model weights are downloaded from Hugging Face Hub (t5-large, gpt2-medium, etc.) — verify checksums independently if required
• No authentication, authorization, or input sanitization for inference — this is expected for a research prototype

If you find a security vulnerability (e.g., a dependency with a known CVE, unsafe deserialization, or path traversal in a script):

1. Do not open a public GitHub issue.
2. Email samyoder23@gmail.com with subject line: [ABI Security] <brief description>.
3. Include: affected file(s), description of the issue, and any known exploit or proof-of-concept.

Expected response time: within 7 days.

This project uses torch, transformers, and sentencepiece. Keep these up to date to avoid known CVEs in those dependencies.