--- title: Your LLM Stack Is a 2008 Database With Better Marketing... description: A senior NVIDIA engineer argues that most production ML security failures aren't exotic AI attacks but classic infrastructure mistakes — misconfigurations,... canonical: https://daily.dev/posts/your-llm-stack-is-a-2008-database-with-better-marketing-lovina-dmello-nvidia-b8by82tvh twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA | daily.dev og:description: A senior NVIDIA engineer argues that most production ML security failures aren't exotic AI attacks but classic infrastructure mistakes — misconfigurations,... og:url: https://daily.dev/posts/your-llm-stack-is-a-2008-database-with-better-marketing-lovina-dmello-nvidia-b8by82tvh og:image: https://api.daily.dev/og/posts/b8BY82TvH.png og:image:alt: Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 20 min read · 1 upvotes · 0 comments ## Summary A senior NVIDIA engineer argues that most production ML security failures aren't exotic AI attacks but classic infrastructure mistakes — misconfigurations, over-privileged accounts, flat networks, and exposed secrets. Drawing on research showing 78% of audited ML setups had critical security flaws, the talk frames LLM stacks as essentially 2008-era databases with new branding. Three core ideas are presented: misconfigurations (not missing features) are the top failure mode; many research-recommended defenses don't survive production due to latency costs; and ML systems must be secured like infrastructure, not just like models. A four-level maturity model mapped to the NIST AI risk framework is introduced, along with practical fixes for the most common misconfigurations. Emerging threats like prompt injection, RAG poisoning, GPU side-channels, and supply chain risks are also flagged. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### why were thousands of Ray clusters exposed on the internet in 2023 Authentication was off by default in Ray, one of the most popular distributed ML frameworks, and teams deploying it to production forgot to turn it on. This left dashboards and job APIs publicly reachable, with security researchers estimating the total exposure at over a billion dollars in accessible compute and data. _Teams tracking ML infrastructure security gaps can follow these lessons on daily.dev as they harden distributed training clusters._ ### what percentage of production machine learning deployments have critical security misconfigurations 78% of 50 real production machine learning setups audited by researchers had at least one critical security mistake, most commonly wide-open access controls, lack of separation between system components, and passwords or trained models sitting in publicly reachable storage. These are basic infrastructure hygiene failures rather than novel AI-specific vulnerabilities. _Engineers deciding how to prioritize ML security work can track findings like this on daily.dev before shipping to production._ ### how much performance overhead does adding security controls to an ML production system cost Basic controls like login checks and input validation cost under about 8% overhead and should always be enabled. Heavier isolation between workloads runs 10-20% and can be applied selectively to sensitive or untrusted workloads. Real-time malicious input detection is the most expensive tier at roughly 15-30% overhead, so it is reserved for higher-risk systems rather than applied universally. _Teams balancing latency budgets against ML security controls can keep up with this tradeoff guidance on daily.dev._ ## Similar posts on daily.dev - [Securing the AI Stack: From Model to Production](https://daily.dev/posts/securing-the-ai-stack-from-model-to-production-auxghvyo4) · InfoQ · 0 upvotes · 0 comments - [we are building data breach machines and nobody cares](https://daily.dev/posts/we-are-building-data-breach-machines-and-nobody-cares-qcskpagwt) · Hacker News · 3 upvotes · 0 comments - [The Complete Guide to LLM Security: Risks, Best Practices, and Solutions](https://daily.dev/posts/the-complete-guide-to-llm-security-risks-best-practices-and-solutions-8ug63q2np) · Orca Security Blog · 0 upvotes · 0 comments - [Schneier on LLM vulnerabilities, agentic AI, and “trusting trust”](https://daily.dev/posts/schneier-on-llm-vulnerabilities-agentic-ai-and-trusting-trust--fkhqbqfs0) · Sutter's Mill · 1 upvotes · 0 comments - [Large Language Model \(LLM\) integration risks for SaaS and enterprise](https://daily.dev/posts/large-language-model-llm-integration-risks-for-saas-and-enterprise-cn1znviyb) · Security Boulevard · 2 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#prompt-injection](https://daily.dev/tags/prompt-injection) [View this post on daily.dev](https://daily.dev/posts/your-llm-stack-is-a-2008-database-with-better-marketing-lovina-dmello-nvidia-b8by82tvh) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA","url":"https://daily.dev/posts/your-llm-stack-is-a-2008-database-with-better-marketing-lovina-dmello-nvidia-b8by82tvh","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/your-llm-stack-is-a-2008-database-with-better-marketing-lovina-dmello-nvidia-b8by82tvh"},"datePublished":"2026-07-20T17:41:33.679Z","dateModified":"2026-09-14T07:03:06.726Z","description":"A senior NVIDIA engineer argues that most production ML security failures aren't exotic AI attacks but classic infrastructure mistakes — misconfigurations,...","image":"https://i.ytimg.com/vi/XjI-AR4pt7Y/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/XjI-AR4pt7Y/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/your-llm-stack-is-a-2008-database-with-better-marketing-lovina-dmello-nvidia-b8by82tvh","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,prompt-injection","timeRequired":"PT20M","video":{"@type":"VideoObject","name":"Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA","description":"A senior NVIDIA engineer argues that most production ML security failures aren't exotic AI attacks but classic infrastructure mistakes — misconfigurations,...","thumbnailUrl":"https://i.ytimg.com/vi/XjI-AR4pt7Y/sddefault.jpg","uploadDate":"2026-07-20T17:41:33.679Z","duration":"PT20M","url":"https://api.daily.dev/r/b8BY82TvH","embedUrl":"https://www.youtube.com/embed/XjI-AR4pt7Y"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/your-llm-stack-is-a-2008-database-with-better-marketing-lovina-dmello-nvidia-b8by82tvh#faq","mainEntity":[{"@type":"Question","name":"why were thousands of Ray clusters exposed on the internet in 2023","acceptedAnswer":{"@type":"Answer","text":"Authentication was off by default in Ray, one of the most popular distributed ML frameworks, and teams deploying it to production forgot to turn it on. This left dashboards and job APIs publicly reachable, with security researchers estimating the total exposure at over a billion dollars in accessible compute and data. Teams tracking ML infrastructure security gaps can follow these lessons on daily.dev as they harden distributed training clusters."}},{"@type":"Question","name":"what percentage of production machine learning deployments have critical security misconfigurations","acceptedAnswer":{"@type":"Answer","text":"78% of 50 real production machine learning setups audited by researchers had at least one critical security mistake, most commonly wide-open access controls, lack of separation between system components, and passwords or trained models sitting in publicly reachable storage. These are basic infrastructure hygiene failures rather than novel AI-specific vulnerabilities. Engineers deciding how to prioritize ML security work can track findings like this on daily.dev before shipping to production."}},{"@type":"Question","name":"how much performance overhead does adding security controls to an ML production system cost","acceptedAnswer":{"@type":"Answer","text":"Basic controls like login checks and input validation cost under about 8% overhead and should always be enabled. Heavier isolation between workloads runs 10-20% and can be applied selectively to sensitive or untrusted workloads. Real-time malicious input detection is the most expensive tier at roughly 15-30% overhead, so it is reserved for higher-risk systems rather than applied universally. Teams balancing latency budgets against ML security controls can keep up with this tradeoff guidance on daily.dev."}}]} ```