New Cheat Sheet: RAG Security#2131
Conversation
|
Hi @mackowski -- just checking if this is ready for review or if there's anything you'd like changed before looking at it. Happy to adjust. Thanks |
|
@razashariff it is ready for review but we need some time to review it ;) |
|
Thanks @mackowski. Just wanted to make sure nothing was blocking on my side. The Secure Coding with AI one (#2132) was also revised based on Jim's feedback. Cheers, Raza |
| - Silently degrade functionality. Users and operators must know when the system is not operating with full security controls. | ||
| - Treat pipeline failures as performance issues. In a security context, a failed retrieval or a failed access control check is a security event. | ||
|
|
||
| ## Section 15: Do's and Don'ts Summary |
There was a problem hiding this comment.
@jmanico what do you think about this? This restates bullets verbatim from Sections 1–14 ans it does not add any new information.
|
Agree with @mackowski on Section 15 : the "Do's and Don'ts Summary" duplicates the per-section Do/Don't lists verbatim and pads the doc by ~40 lines without adding new content. The "Implementation Priority" block at the top already gives readers a quick-reference. Recommend dropping Section 15 entirely, or replacing it with a one-line-per-control checklist (no prose) so it serves as a true at-a-glance reference rather than a restatement. Minor (optional):
Otherwise solid : the access-control inheritance, fail-closed, and tool-invocation sections are particularly well done. LGTM after Section 15 cleanup. |
|
Hey @jmanico @mackowski -- thanks for the thorough review, really appreciated. All feedback addressed:
Let me know if anything else needs tightening. Thanks again, cheers Raza :) |
|
Thanks @razashariff |
3 participants
Addresses #2126.
Thanks @mackowski for approving and assigning this one — here it is.
RAG is now standard architecture for enterprise AI but introduces a unique attack surface distinct from both traditional web vulnerabilities and standalone LLM risks. This covers the topic practically with a first-approach focus that I hope will help our community immensely.
14 sections covering the complete pipeline from document ingestion through to output validation:
Each section has practical Do/Don't guidance. Happy to iterate on feedback.
I have also built DVRAG (Damn Vulnerable RAG Pipeline) as a companion training tool — a deliberately insecure RAG system with 25 vulnerabilities mapped to each section of this cheat sheet. Practitioners can use it to understand and test each attack vector hands-on. Happy to share details if useful for the community.
Thanks again — Raza Sharif :)