# Securing AI and LLM Workflows: Key Takeaways from our webinar with Derek Manning from Superna **URL:** \ **Category:** Use Cases\ **Created:** [June 4, 2025, 3:59pm UTC](https://community.vastdata.com/t/securing-ai-and-llm-workflows-key-takeaways-from-our-webinar-with-derek-manning-from-superna/1416 "2025-06-04T15:59:12Z")\ **Posts on this page:** 1\ **Page:** 1
**Author:** ![jonasrosland](https://sea2.discourse-cdn.com/flex022/user_avatar/community.vastdata.com/jonasrosland/32/1_2.png) [@jonasrosland](https://community.vastdata.com/u/jonasrosland)\ **Post date:** [June 4, 2025, 3:59pm UTC](https://community.vastdata.com/t/securing-ai-and-llm-workflows-key-takeaways-from-our-webinar-with-derek-manning-from-superna/1416/1 "2025-06-04T15:59:12Z")
Hey folks! 👋 Just wanted to share some highlights from yesterday’s community webinar with Derek Manning from [Superna](https://superna.io/). If you missed it, don’t worry - you can watch it below on YouTube, but here are the key points that really stood out to me. [![](https://us1.discourse-cdn.com/flex022/uploads/thecosmosai/original/2X/3/3ca7abfbda01d230048dea1eaffcdf879b719673.jpeg "Securing AI & LLMs: How Superna and VAST Protect the Data That Powers Innovation") ](https://www.youtube.com/watch?v=haeaRqNdOlY) ## The Reality Check: AI Security is a Data Problem Derek started with something that really hit home - **99% of Fortune 500 companies now rely on data-driven intelligence to stay competitive**. It’s not just an advantage anymore; it’s table stakes. But here’s the kicker: most organizations are focusing on the flashy AI/ML parts while leaving their data foundations vulnerable. The webinar walked through what Derek calls the “data-driven intelligence workflow”: - **Clean** → Remove PII, ensure compliance - **Pool** → Centralize data (hello, VAST Data customers!) - **Integrate** → Merge with external sources - **Enrich** → Refine and enhance datasets - **Analyze** → Extract insights and train models ## The ByteDance Wake-Up Call One story that really drove the point home was the [ByteDance incident](https://arstechnica.com/tech-policy/2024/10/bytedance-intern-fired-for-planting-malicious-code-in-ai-models/) from October 2024. An intern allegedly planted malicious data in their AI training pool, potentially causing **$10M+ in damage**. While ByteDance said the figures were exaggerated, it shows how vulnerable these workflows can be - and that’s just what we know about publicly! ## Enter “Cyberstorage” - Gartner’s New Category This was news to me: Gartner has created an entirely new security category called **“[cyberstorage](https://superna.io/gartner-2024-strategic-roadmap-for-storage)”**. Traditional security stops at the network/endpoint level, but attackers go straight for the storage once they’re in. Makes total sense when you think about it. ## How Superna + VAST Work Together The integration is pretty darn neat: - **Real-time audit stream monitoring** via integrations with VMS - **API-based relationship** with VAST for automated snapshots and user lockouts - **User behavioral analysis** to detect abnormal patterns - **Compelling event-based snapshots** (not just scheduled ones) - **Zero-trust API** integration with security tools like ServiceNow What I found really cool was the “honeypot tripwire” approach - creating non-production shares that trigger alerts if accessed. Simple but very effective. ## The “Left of the Boom” Philosophy Derek explained ransomware attacks using a timeline with a “boom” in the middle. Traditional detection often happens on the right side (users calling about encrypted files, pure nightmare fuel), but Superna focuses on the left side - catching attacks before and as they happen, not after the damage is done. ![Left of the Boom](https://us1.discourse-cdn.com/flex022/uploads/thecosmosai/original/2X/d/d59f870456ba37bd35880bd0ed51e625176ec5a8.jpeg) Their response is automated: 1. Detect abnormal behavior 2. Create compelling event snapshot 3. Lock out suspicious user 4. Alert security teams via webhooks 5. Surgical recovery of affected files ## Key points from the Q&A - **Data cleaning challenges** : PII removal is the biggest hurdle for AI compliance - **Tools** : There are many tools for identifying and cleaning data, everything from license plate/credit card detection to medical image depersonalization - **Data selection** : When looking at what data to include in AI training, try to include as much as possible - emails, CRM data, historical reports. The “unknown unknowns” in your data might be the most valuable ## Bottom Line If you’re running AI workloads on VAST, this strategic partnership makes a lot of sense. The security landscape is changing fast, and having data-layer protection that works natively with your storage platform seems like a no-brainer. Thoughts on the cyberstorage concept? Would love to hear how others are thinking about securing their AI pipelines.