Cyber Risks · Trail of Bits Blog

Auditing the Ask Astro LLM Q&A app

Summary: Trail of Bits audited Ask Astro, an open-source RAG chatbot built on A16Z's reference architecture, and found four issues including high-severity vector database poisoning via front-running, where an attacker posts a forum thread before ingestion and deletes it afterward. Other findings include low-severity split-view poisoning through GitHub issue ingestion bugs (remediated in PR #325) and a medium-severity Weaviate client flaw in the _sanitize_str function that could leak non-public documents if the vector database shares infrastructure with a private one.

What's compelling: Ask Astro is a reference implementation many new RAG developers will copy, and its ingestion pipeline treats community forum posts and GitHub comments as no less authoritative than official documentation - so an attacker who knows the ingestion interval can inject arbitrary text into the knowledge base and delete the evidence before moderators see it.

Read the original at Trail of Bits Blog ↗

Topics
AI Security
Related