AI Showdown: The Good, The Glitchy and The Risky

Nowadays, with bad actors trying to compromise open-source projects "en mass", it’s important to put aside the excitement about AI just long enough to make the right decisions. Introducing AI into your CI/CD system might sound fun but there are reasons why making the wrong choices will bite you and your organisation.

In this session, we’ll look at each of the top AI frameworks and assess them under the cold light of day. We’ll look at their background, how (and on what) they are trained etc and generally try to assess accuracy and usefulness today and tomorrow.

We’ll also look at which ones have good security, which ones follow great coding practices and even which ones pass the emerging open-source standards and legislation being developed for components and AI models by organisations like OWASP, OpenSSF and governments worldwide.

In AI, the best framework isn't just smart, it's the one that doesn't outsmart your safety.

Transcript

1. 🤖AI Showdown: The Good, The Glitchy and The Risky JNation 2024, Olimpiu Pop
2. All glitchy visuals are courtesy of Steve Poole https://extrapolate.app JavaAdvent.com
3. Part I: Debunking your POV: “The Good, The Bad and The Glitchy” History 🥱 + Challenges A (Gen) Picture Is Worth 1K Examples Part II: AI Showdowns AI for the greater good AI usage by the bad guys AI goes awry Part III: Generative AI and the Developer Gen AI in SDLC Consequences to Software Supply Chains/ Potential Legal Issues Gen AI Reference Architecture Part IV: Gen AI Puzzle Pieces: Where to get started? Cloud Native + AI = ❤ Safety Frameworks (Governance 🤫) Java Landscape
9. "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." — Edsger W. Dijkstra
10. “Create a visually striking image that captures a philosophical inquiry, divided into two distinct halves. The top half features a sleek, modern computer or AI system positioned against a backdrop that suggests intellectual endeavor, complete with a thought bubble that includes symbols of thought or intelligence like a light bulb or a brain. The bottom half showcases a submarine gliding through the ocean's depths, with its own thought bubble containing symbols of swimming, such as swim fins or a swimmer's silhouette. This composition aims to provoke thought on the nature of artificial intelligence and technology, mirroring the quote "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." Use a blend of modern and whimsical elements to highlight the juxtaposition and provoke contemplation on the capabilities of technology versus natural action.” Create image to support "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." Human ChatGPT image generator plugin Dall-E Image
11. Gen AI = Generative AI [ ʤɛnərətɪv eɪ-aɪ Generative AI can be thought of as a machine-learning model that is trained to create new data, rather than making a prediction about a specific dataset. A generative AI system is one that learns to generate more objects that look like the data it was trained on ] 11 https://news.mit.edu/2023/explained-generative-ai-1109
12. The big breakthrough: textual understanding and digital content 🤷: describe this image…please! 🤖: … The scene is reminiscent of a classic movie scene from "E.T. the Extra-Terrestrial," where a child bicycles across the moon's face with an alien in his basket …
13. 󰤈: Describe the content of the first image in more detail 🤖: “200 bla bla blabla bla bla bla bla bla bla bla bla(s) …” 🤖󰥚 : Create this image … The big breakthrough: textual understanding and digital content
14. Original Dall-E DA VINCI ADOBE Scene from the film ET where ET and Elliot flew in front of the moon on a bike I'm unable to directly reference copyrighted material or recreate specific scenes from movies like "E.T. the Extra-Terrestrial."
15. AI Image transformation is amazing - but always glitchy Turn this image into one where the person is in a spacesuit Turn them to face right Remove the helmet Turn them slightly to their right
16. Place him on a spaceship
17. Thoughts On Generative AI 1: They are incredible creative mapping tools 2: They are hard to control exactly - the mapping is lossy 3: Each GenAI tool will produce different mappings. Even with the same prompt 4: Every generation has glitches - where the mapping fails or goes in unexpected directions 5: Your choice of AI will lock you in. 6: Gen AI’s are only as good as the data they have access to and the human supervision. 7: Every solution with Gen AI in it is going to have multiple AI elements - including more traditional discriminative AI tools. Your software supply chain will be complex.
18. II: Showdowns
23. Part 3: Gen AI and The Developer
26. IDEs didn’t scare us, AI shouldn’t either?! https://hollycummins.com/will-ai-take-our-jobs/
28. https://danielmiessler.com/p/the-ai-attack-surface-map-v1-0/
30. https://arstechnica.com/information-technology/2024/01/ai-poisoning-could-turn-open-models-into-destructive-s leeper-agents-says-anthropic/ Poisoned models
31. Poisoned models
32. Software Supply Chain today
33. Software Supply Chain tomorrow
34. Gen AI Adoption In Dependencies Jobs will be taken from developers by developers (AI augmented) https://www.sonatype.com/state-of-the-software-supply-chain
35. (more) Thoughts On Generative AI 8: Unprovenanced Generative AI models are inherently dangerous 9: Adding Gen AI into your supply chain is much more complicated than you might expect. 10: Java Devs are not used to the types of security problems that occur 11: Exposing ChatGPT style bots to the public is likely to be problematic 12: Securing AI is a brand new, scary topic. DevSecAISecOps anyone? 13: Legislation on AI usage is well underway. Make sure you’re up-to-date - implemented in waves 14: Say hello to your data scientists - it’ll be your job to teach them Software Engineering and security
37. IV: Picking Up The Tools!?
42. March 2024
43. CNCF: Cloud Native Reference Architecture https://www.cncf.io/wp-content/uploads/2024/03/cloud_native_ai24_031424a-2.pdf
44. Open Source Security and Risk Analysis Report, Synopsys - 2021: By some estimates* 84% of all codebases have at least one vulnerability, with an average of 158 per codebase. The majority have been in the code for more than 2 years and have documented solutions available. It might be too windy…
45. Part of the Solution: OpenSSF ScoreCard Malicious maintainers Source code compromises Build system compromises Malicious packages
46. What Options Does The Java World Provide? Langchain4j quarkus-langchain Spring-AI JLama
47. Langchain4j (3.6k ⭐) • Vanilla Java • No framework dependency • Advanced RAG Practices • Multiple models support • Multiple vector DBs support • Rapid release cycle - 0.29 • Fluent API • Hard to keep pace with the mother ship (langchain) Score 6.4/10 • Contributors + Maintained • No dangerous workflows • License + Security Policy • Packaging • No Vulnerabilities • No SAST • No Pinned Dependencies • No Fuzzing • No Dependency Update Tool • CI Tests • Branch-Protection Not Enable • CodeReview https://github.com/langchain4j/langchain4j
48. Quarkus Langchain (109 ⭐) • The cloud-native promise • Declarative • RAG Tooling langchain • Rapid release cycle • The last link in the chain • Contributors + Maintained • Dangerous workflows • License + Security Policy • Packaging • No Vulnerabilities • Code Review • No SAST • Pinned Dependencies • No Fuzzing • No Dependency Update Tool • No CI Tests • Branch-Protection Not Enabled Score 5.7/10 https://github.com/quarkiverse/quarkus-langchain4j
49. Spring AI (2.3K ⭐) • Spring Ecosystem • Spring-Way of Doing things • Default OpenAI + Azure OpenAI • Built from scratch inspired by LlamaIndex, Langchain • Contributors + Maintained • Dangerous workflows • License + Security Policy • Packaging • No Vulnerabilities • No SAST • Pinned Dependencies • No Fuzzing • No Dependency Update Tool • No CI Tests • Branch-Protection Not Enabled • CodeReview lacking Score 5.0/10 https://github.com/spring-projects/spring-ai
50. JLama (338 ⭐) • Vector API • Fallback to Panama • From scratch implementation • Contributors + Maintained • Dangerous workflows • License + Security Policy • Packaging • No Vulnerabilities • No SAST • Pinned Dependencies • No Fuzzing • No Dependency Update Tool • No CI Tests • Branch-Protection Not Enabled • CodeReview lacking Score 5.7/10 https://github.com/tjake/Jlama
51. Playing Around • Ollama - “Docker Of LLM” • Podman - “Closing the Gap Between PoC and Prod” • JLama - “First Java Inference Library in MavenCentral” • Devoxx Genie - “Visual Tool for Comparing Models”
54. • Popularity != Safety (Use OpenSSF Scorecard to choose the best tool) • Read the licenses - they might be washed • Limit the blast radius by containing them • Keep the humans in the loop • Continuously test the outcomes • Find the proper size for your models, you seldomly need L generative models (smaller == smaller attack surface) • The risk is higher as LLMs are exposed to both social engineering and cyber tactics Thoughts…again…
55. Thank you