MAVOCO’s Post

Day 7 and 8 of contributing to open source❤️ until I find a remote job 💻 Continuing with OWASP® Foundation's Top 10 ML project, I finished working on writing summaries for all the attacks and vulnerabilities present in the project. The inspiration for this PR: https://lnkd.in/dfwJimVv has been drawn from OWASP's Large Language Model project: https://lnkd.in/dGRkYv5V Summarisation has always been crucial for any project where details can be summed up in a short paragraph. I have previously worked with summarisation at a startup as an NLP Engineer and I understand how important it is to have this section in modern projects. Soon, I will be taking up issues from the repository and solving them one by one which I have been planning since a long time. #buildinpublic #github #linkedin #machinelearning #nlp #remotejob #gitlab #opensourcecontributions #python #pythonprogramming #documentation #opensource #owasp

I recently stumbled upon a GitHub repository that caught my eye – a cutting-edge application of #GenAI. At first glance, I was like, "Wow! This is a niche use-case for GenAI!" But as I pause, I found myself asking, "Is this right? Are we truly prepared for the risks involved?" If you're considering implementing this in a production environment, it's worth pausing to reflect. Beyond your technical capabilities, do you have the necessary time and expertise to look under the hood of such a tool? We must not overlook the #SecurityImplications, which include potential vulnerabilities in the generated code or dependencies. Nor can we ignore the regulatory and compliance factors, particularly in industries subject to stringent regulations. #Threatmodeling should be a key part of this evaluation. And let's not forget the #DataPrivacy concerns that come with handling sensitive information. There seems to be a trend in bringing ideas to life faster through automation, and it's certainly a welcome #innovation. However, we must #exercisecaution. As appealing as these cutting-edge solutions may be, we need to approach them with a healthy dose of skepticism and due diligence. After all, innovation should never come at the expense of security, compliance, and trust. Let's embrace technology but not without a robust plan that includes #ThreatModeling, risk assessment, and a clear understanding of legal implications. #GenAI #InnovationWithCaution #RegulatoryCompliance #SecurityInTech

Unraveling the CIA Kryptos Code (Part 2) - Yogesh Jadhav delves into the complexities of transposition ciphers, exploring the intricacies of keyed columnar transposition and route transposition. With detailed examples and TypeScript code, this post provides a comprehensive understanding of these encryption techniques. Stay tuned for the next post where we decode Kryptos 3. #Cryptography #TranspositionCipher #NSA #KryptosSculpture #AI #ML #DataScience

LangChain allows to connect LLMs with external tools and data sources. The LLM converts the user input into a format supported by a specific tool (for example text to json or text to SQL) and uses LangChain framework to call the tool. While it is truly awesome and allows amazing applications, be awere, that historically parsers and deserializers which took 3rd party data as input are the weakest links in security. Remember log4j? In essence it is a parser/deserializer (in logger) which allowed to execute external java class, but something went wrong (of course). Natural language input much harder to predict and test in comparison with structured input. Will we see interesting variations of such vulnerabilities in systems using LangChain-like agent-based frameworks in future? #ai #agents #vulnerability #artificialntelligence

AI is awesome...but it is not perfect and has the potential to expand weaknesses in your codebase. The cost can be in the millions as this leads to breaches, outages, and more. And to compound this risk, today, audits often only cover a minor percentage of code changes, resulting in the the potential for exploits staying undetected. Use Change Captain to help offset the risk of accidental or malicious code changes that occur without proper review and approval. Drop us a line ⚓ and get onboarded today. https://lnkd.in/e2Ffq2hU

Question Answering (QA) over documents is one of the most common use cases for LLM applications. However sending our private and confidential documents over to an LLM API provider like OpenAI is not a great idea. Why not do it all locally on your laptop with Llama 2? I wrote an article showing how it can be done. https://lnkd.in/gXSpZDRn

🤯 The framework in LLM Space is growing - And there are now lots of startups in this field - #PromptEngineering, #RAG, #LLMOps, #LLMSecurity, #Fairness, #Guardrails 📈 When we think about the tools, the challenge grows exponentially 🤔 If you don't know where to start, well... you're not alone. 🤙🏻 The good thing is - we untangle these topics in the newest LLMs.HowTo post 🔥 In this post, I have published a blog on my friend's 🚀 Rafael V. Pierre 's great website. Please do visit https://llmshowto.com/ to know on the following topics 👩🏻💻 Prompt Engineering 🔎 Retrieval Augmented Generation (RAG) 🧪 LLM Evaluation 👮🏽♀️ LLM Security 🏭 LLMOps 🛠️ Which tools are interesting to keep an eye on ...and more. Please subscribe my newsletter https://lnkd.in/eSxBSav8

News A look at the current state of API specification languages, in particular looking at TypeSpec. https://lnkd.in/dGebEx36 A look at using VolumeSnapshots in Kubernetes to speed up pod startup time for certain types of application. https://lnkd.in/dNMr53FE A look at the pros and cons of using JSON for structured logs. https://lnkd.in/d6umCvDr A nice walkthrough of building a CI/CD pipeline using GitHub Actions, with a focus on meeting SOC 2 compliance rules. https://lnkd.in/dnpSDr_D

Interesting idea on Unsupervised Evaluation of Code LLM: Roundtrip Correctness (RTC) For Code Synthesis - Evaluate LLMs by giving them a piece of code, having them describe it, and then asking the LLM to rewrite that code given only the description. The metric is the similarity between the original code and the rewritten code. https://lnkd.in/eHpV6Y5q