- OWASP Top 10 LLM Apps Newsletter
- Posts
- OWASP LLM Top 10 Newsletter - May '24 Edition
OWASP LLM Top 10 Newsletter - May '24 Edition
V2 discussions, relevant news and links from the community
Krishna Sankar
May 21, 2024
Steve Wilson’s rendering of V2 efforts !
Greetings Gen AI Security Enthusiasts and OWASP Community Members!
Musings from the editorial desk.
The first challenge, as a newsletter editor, is to select the most relevant subset of materials - especially when it is about Generative AI where there is no lack of blogs, announcements, factuals & counterfactuals.
For the readers, the newsletter needs to be insightful, manageable and respectful (of their time) by not repeating (to much) from all the other newsletters.
The second challenge, of course, is to present a succinct summary - what to leave out is more important than what to leave in !
I hope I am succeeding in this Sisyphean task - with the boulder at the bottom of the Tartarus hill 😇, every month
The focus for this month is the OWASP LLM Top 10 v2. Time for lots of preparation, thinking, proposals, deliberation and finalization.
I am sharing couple of my ideas here as a start … Steve already has a few entries
And, … we have the highlights from the v1.1 survey - very illuminating and in tune with the fast pace of Generative AI
The June Edition will focus on RSA 2024 with:
The highlights, summary and notes from our RSA conference sessions to give you a quick overview.
Additionally, we will include the links, so that you can watch the sessions at your convenience for a detailed study
General Note : Please keep an eye on the upcoming OWASP Global AppSec conferences @ Lisbon & San Francisco - Attend, propose training, submit papers and share with your social media presence,…
I have followed my own advise and have submitted a two day training “LLM Security Analysis & AI Red Teaming Workshop“ covering Principles, Policy (OWASP LLM Top 10, NIST RMF et al), Practice & Governance of GRT (Generative AI Red Teaming) with a 70-30, mix of hands-on (using good and uncensored LLMs) vs theory, form factor. We will see if it makes the cut …
📋 Before we dive into the v2 discussions …📋
An interesting interview by our own Sandy Dunn “Inside the OWASP Top 10 for LLM Applications” - The inimitable duo - Mike Shema and John Kinsella -interviews Sandy Dunn, at the Application Security Weekly !
The show notes has relevant links
“Just how large a list does a large language models largely needs” - Of course Sandy has an eloquent answer …
And, …. the “how large” is the looming question for the v2 discussion ! Very timely … keep on reading on how you can influence the evolution of this largely large, large language model list ! …
Mike addresses the “packets, protocols and programs” with an elegant limerick - “I am more than a math computation; An LLM said with elation; It sounded like magic; But, then slightly tragic; Being just a hallucination” (Pl note the AABBA rhyme scheme!)
Asymptotical Detour - Notice the Dune book in the background - I am a big fan of the Dune universe. In fact, have a blog coming up on All things Dune - The music (Hans Zimmer/Dune Sketchbook/Song Of Sisters), the lego (Dune Atreides Royal Ornithopter - bidding for one at eBay as we speak!), the books (36 of them, incl the prequels & sequels) and of course, the movies.
Sandy started with a succinct summary of the origin of the OWASP Top 10 LLM. Very interesting to hear about the origins !
Then they continued to insightful discussions on the dual nature of AI, training, data, the large attack surface, the challenges enterprises are facing in terms of guarding LLM based applications, the focus and the audience for the Top 10 LLM list
Interestingly Mike was asking about #s 11,12,13 and 14.
I have an answer - a list that I am proposing for v2. More later below
I really liked Sandy’s insights on many of the conundrums we face in terms of Generative AI - for example we put human values in the evaluation of LLMs which is not so deterministic, hard and fast. The nuances are too finer to be internalized by an LLM.
A closely related interesting point in my blog here.
LLMs are after all smooth talking Stochastic Parrots !
Mike asked about the evolution of the list as well as the feedback.
Very timely question as we had the 1.1 survey out few weeks ago and have the v2 effort that we will talk about next !
At the end, Mike asks Sandy the customary quiz - “Appsec is most likely to …” and Sandy’s completion “… be more critical than it’s ever been before” ! Well said, Sandy. Very thoughtful …
In fact, the LLM security/safety/trust should get progressively more stricter as LLMs become more powerful - i.e., the rate of increase in the demonstrated confidence (in LLMs) should exceed the rate of increase in LLM power !
🔟 OWASP Top 10 LLM 2.0 ! 2️⃣
Steve says it all ! The New Entry Submission link here
Thanks for all for completing the v1.1 survey (You did complete the survey, didn’t you ?)
Steve has shared some interesting results here. Slide #11 - v2.0 Major Targets gives a good summary.
As an example of v2 proposal entries, I plan to submit this list as candidates (the #11, #12, #13 & #14 Mike was asking about) for discussions, addressing three areas viz:
Cover Security (of the operator), Safety (of the user) and Trust (by the user) of LLM Applications
Be the one-stop-shop to answer the question "What are the top risks that I should worry about when I deploy my application that has LLM components?"
We should make the artifacts approachable, accessible and consumable by a wide variety of audience - folks who have high information overload and low attention span (current company included !)
A good opportunity to submit yours …
My proposed list to make the Large Language Model list larger
📺 Heard on the ‘Net 📺
OWASP Switzerland Summit: Recharge & Reconnect - “After more than two years with only online talks, we are going physical again”
The OWASP Foundation Celebrates 20th Anniversary ! Actually the OWASP community is 23 years old
OWASP LA in-person meeting on 5.22.24 - they are hosting our own Steve Wilson !
And,… don’t miss the APISEC.CON 2024 to get a glimpse of Tony Stark and our own Aubrey King !
🔗 OWASP LLM Documents & Links 🔗
📱Stay Connected📱
Aubrey King has expanded our social media reach across various platforms including LinkedIn, Youtube, Twitter (Agree with Will, not going to say X), and soon even more! Follow and don’t be shy to re-share any content we post! BTW, if you're presenting anything related to our project externally, do let us know at [email protected] so we can review and blast out to our social media followers!
💭 Closing Thoughts 💭
OWASP Top 10 LLM - V2 work is upon us. The world of LLMs have changed since the v1. Please submit your favorite items … We look forward to seeing you in our Slack channels and at our upcoming meetings!
Till next time … Stay Secure & Stay Sharp …