Securing AI Operations with Agent Sandboxes

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Securing AI Operations with Agent Sandboxes · Channel: VAST Data · 03:44 · Watch on YouTube

Topics: AI agents, Data, databases and embeddings, AI infrastructure and hardware, AI safety

Summary

Agents can multiply the productivity gains of frontier models, but organizations need to give them identities, memory, secure sandboxes, observability, runtimes, and safe access to data. VAST describes using its data store, data engine, database, Insight Engine, and agent engine to provide these capabilities, including shared sandboxes, data discovery, and monitoring. Low-latency access to tabular data can help agents work faster, while a distributed, multi-tenant platform aims to simplify deployment at scale. VAST says it will work with customers to prepare their data for AI and bring agents into production.

Key points

Agents can multiply individual productivity gains by running on people’s behalf.
Agents need identities to limit access and trace actions, along with memory, secure sandboxes, and observability.
Agent runtimes need to support execution, sleep and wake cycles, and potentially migration between machines after failure.
Agents need secure access to structured and unstructured data to perform useful tasks.
VAST describes its data store as a shared file system where agents can create snapshots, revert to earlier states, and clone data for subagents.
The VAST data engine processes changing or newly ingested data, while the database stores semantic representations using a vector index.
Insight Engine provides an embedded RAG pipeline as an MCP tool for finding relevant text and media.
Agent telemetry can support monitoring, auditing, optimization, fine-tuning, and reinforcement learning.
VAST says its distributed, multi-tenant architecture can provide storage and services without provisioning hardware and software separately for each tenant.
The video highlights agent startup and operating cost, and task completion speed, as important metrics.

Chapters

Start	Chapter	Summary
00:00	Why agents need secure foundations	The speaker explains agents’ productivity potential and the need for identity, memory, sandboxes, observability, runtimes, and data access.
01:03	VAST data services for agents	The video outlines how VAST’s data store, data engine, database, and Insight Engine support agent collaboration and data discovery.
02:09	Telemetry, runtime, and low-latency data	The speaker describes monitoring and optimizing agents, the agent engine, and why fast access to tabular data matters.
03:13	Scaling agent deployments	The video presents VAST’s distributed multi-tenant platform, identifies cost and speed as key metrics, and describes plans to help customers bring agents to production.

Quotes

Unlocking agents is the one of the biggest productivity multipliers organizations are chasing. — 00:00

agents need to run in a secure sandbox to protect the organization from agents going rogue. — 00:31

They can create snapshots, travel back in time if they made a mistake — 01:35

they’re as fast as their slowest tool. — 02:40

Because agents make everything fast. — 03:44

Concepts and entities

agent sandbox, VAST Data, Vast data store, Vast data engine, Vast database, Insight Engine, MCP, text-to-SQL, multi-tenancy

Full transcript (original language)

00:00 Today, I want to talk about agents. Why are they important? Unlocking agents is the one of the biggest productivity multipliers organizations are chasing. We all felt the productivity boost of having access to frontier models as individuals, but agents multiply that. They multiply it by the number of agents running on our behalf. So, how does one enable agents, right? Agents are similar to humans in some ways, but eventually they’re computer programs. They need an identity like a human has [music] to limit their access and for us

00:31 to trace their actions. They need memory like humans have, but unlike humans that are accountable for their actions, agents need to run in a secure sandbox to protect the organization from agents going rogue. And we need the observability to find out when and how and why it happened. Now, because they’re computer program, they need the runtime, an environment or sandbox where they can execute, where they [music] can go to sleep, wake up, and potentially even migrate across machines on failure. Lastly, if we put them in a cage, being

01:03 completely isolated from the world, they’re not very useful. They need tools to access data to perform actions. And accessing data is the biggest challenge and opportunity as it has to be done securely and support many different shapes of data. Unstructured data like text, images, videos, audio clips, but also structured data sitting in tables where a lot of the enterprises business data is currently stored. So, what is Vast doing to enable agentic AI? The Vast data store is now being used for agent sandboxes. It’s your familiar

01:35 file system, but agents collaborate on it. They can create snapshots, travel back in time if they made a mistake, and they can clone it to allow sub agents to work off of the same data. The Vast data engine is our serverless compute framework that continuously processes data as it’s changing or being ingested into the system. The Vast database then holds a semantic representation of all of your data leveraging a new trillion scale vector index. And agents then use Insight Engine, which is the our embedded rag pipeline, as an MCP tool to find text and media

02:09 that’s relevant to its task. Now, another great use case for the database is storing telemetry to enable monitoring, auditing, and optimization loops. The same principle applies to agents where their past experience can lead to optimizations through fine-tuning, reinforcement learning, or simply optimizing the prompts. The agent engine is the runtime supporting the agent with a sandbox, the tools, and it makes deployment easy and scalable. What we’re talking about today is the ability for an agent to converse with tabular data at low latency. The

02:40 conversational part of an analytics pipeline would be familiar to people through what’s called text-to-SQL. But why does low latency matter? Remember, agents are not humans. They’re faster than us, but they’re as fast as their slowest tool. And this is where efficient and scalable data structures translate into faster agents, better hardware utilization, and therefore just better outcomes. So why VAST and why now? Deploying agents is like deploying any other application. Setting up all of the above components as discrete solutions, storage, databases, monitoring, and

03:13 search makes deploying AI complex and slow. Doing this at scale is a next-level challenge that rarely was tackled in this cloud era. It’s become more difficult [music] to tackle at that agent scale. And this is why we’ve built a new distributed system architecture that has scale and multi-tenancy as core tenants of the platform. Our universal storage platform enabled organizations to deploy one fungible pool of flash to serve different tenants and workloads, while isolating their data and resources. As our platform evolved into an AI operating system, we’re now able to

03:44 deliver millions of file systems, buckets, event streams, databases, without running additional services, without provisioning hardware and software on a per tenant basis. Now, this also sets focus on new metrics that matter most in this agentic world. What is the cost of starting and running an agent and the services that comes with it? And how fast does an agent work to accomplish its task? So, what’s next? We’re going to be working with our customers on making their data AI ready and bringing agents to production. Because agents make everything fast.