--- title: OpenAI Agents SDK update adds sandboxed workspaces and... description: A survey of the current state of AI agent infrastructure, governance, and security threats heading into 2026. OpenAI's Agents SDK now supports sandboxed... canonical: https://daily.dev/posts/openai-agents-sdk-update-adds-sandboxed-workspaces-and-cloud-storage-support-hyrsyxgpz twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: OpenAI Agents SDK update adds sandboxed workspaces and cloud storage support | daily.dev og:description: A survey of the current state of AI agent infrastructure, governance, and security threats heading into 2026. OpenAI's Agents SDK now supports sandboxed... og:url: https://daily.dev/posts/openai-agents-sdk-update-adds-sandboxed-workspaces-and-cloud-storage-support-hyrsyxgpz og:image: https://api.daily.dev/og/posts/HYRsYxGpz.png og:image:alt: OpenAI Agents SDK update adds sandboxed workspaces and cloud storage support 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. # OpenAI Agents SDK update adds sandboxed workspaces and cloud storage support **[Collections](https://daily.dev/sources/collections)** · 5 min read · 3 upvotes · 0 comments ## Summary A survey of the current state of AI agent infrastructure, governance, and security threats heading into 2026. OpenAI's Agents SDK now supports sandboxed workspaces, container snapshotting, and 100+ LLMs, while deprecating the Assistants API in mid-2026. Cloudflare moved its Sandboxes and Containers to GA with features like secure credential injection, PTY terminal support, and snapshot-based session recovery cutting warm starts from 30s to ~2s. Despite rapid tooling maturity, governance lags badly: only 21% of companies have mature governance for autonomous agents, and 60% can't terminate a misbehaving agent. Key failure modes include missing agent registries, permission creep, prompt injection, and cascading failures in multi-agent chains. On the threat side, attackers are already deploying coordinated agent swarms for fraud — autonomous identity generation, account creation, and cashout — operating at the interaction layer where most security stacks are blind. ## Content ## What's actually shipping The past few months have seen a wave of concrete infrastructure releases aimed at making AI agents production-ready — not just demos, but the plumbing underneath them. **OpenAI's Agents SDK** got a significant update. The headline addition is sandboxed workspaces that separate the agent harness from the compute it runs on. Developers can bring their own containers or use partners like Cloudflare, E2B, Modal, Blaxel, and Vercel. Agents can mount cloud storage from S3, GCS, Azure Blob, or Cloudflare R2, and the SDK supports container snapshotting for stateful workflows. The harness itself brings Codex-style scaffolding — instructions, tools, approvals, tracing, handoffs, resume bookkeeping — to the broader SDK, launching in Python first with TypeScript to follow. The SDK supports 100+ LLMs via the Chat Completions API. One housekeeping note: the Assistants API is being deprecated with a mid-2026 sunset, so teams should start migrating to the Responses API now. **Cloudflare** had a busy Agents Week. Sandboxes and Containers moved to general availability — persistent isolated Linux environments built for agent workloads. New features include secure credential injection via an egress proxy (agents never see the actual tokens), PTY terminal support over WebSocket, persistent code interpreters that maintain state across calls, filesystem watching via inotify, and snapshot-based session recovery that cuts warm start time from 30 seconds to about 2. Pricing is now active CPU billing at $0.00002 per vCPU-second. Figma is already running production workloads on it. Beyond sandboxes, Cloudflare also shipped: Agents SDK v2 preview, a voice pipeline, Email Service beta, unified inference across 14+ providers, Agent Memory, AI Search, Browser Run with Live View, a new unified CLI, and an Agent Readiness score for the broader web. It's a lot. The through-line is positioning Cloudflare as the infrastructure layer for agents the way it became the infrastructure layer for the web. Cloudflare also shared how they built their own internal AI stack on their own products. The numbers are interesting: AI Gateway handles 20M+ requests/month and 241B tokens, a security agent runs 7B tokens/day on Workers AI at 77% lower cost than proprietary models, and merge request volume nearly doubled quarter-over-quarter after rolling out AI coding tools across 3,900 repos. --- ## The governance gap nobody talks about until something breaks Here's the uncomfortable part. The tooling is maturing fast. The governance isn't. A Deloitte/MIT survey found 74% of companies plan to deploy agentic AI within two years, but only 21% have mature governance for autonomous agents. Dynatrace research shows 50% of enterprises already have agentic AI in limited production. And according to Portkey's data, only 14.4% of agents go live with full security approval, and 60% of organizations can't terminate a misbehaving agent when something goes wrong. That last number is worth sitting with. You can't kill it. It's just running. The failure modes that emerge at scale are predictable in retrospect but easy to miss when you're deploying your first few agents: - **No registry.** Nobody knows which agents are running or who owns them. - **Permission creep.** Agents accumulate access beyond their original scope because nobody audited it. - **Prompt injection.** Any agent that reads untrusted content is a potential attack surface. - **Cascading failures.** Multi-agent chains fail in ways that are nearly impossible to debug without full traces. The argument from people who've hit these walls is that governance — registries, governed endpoints, access controls, observability, kill switches — is the real bottleneck at scale, not model performance. The models are good enough. The operational infrastructure around them isn't. Non-human identities already outnumber human ones in most enterprise environments. Agentic AI will accelerate that gap considerably. The recommendation from security teams is a centralized control plane that governs agent permissions, policies, models, and tools — something that enables auditability before you need it, not after an incident. --- ## The fraud problem is already here While enterprises are still figuring out internal governance, attackers have already figured out how to weaponize agents externally. The current threat model looks like this: a single operator deploys 20+ specialized agents that execute a full attack chain — synthetic identity generation, account creation, credit history building, coordinated cashout — with no human effort per account. It's essentially a digital factory running autonomously. What makes this categorically different from traditional bot fraud: 1. **Autonomous iteration.** These agents probe defenses and adapt. They don't just retry the same request. 2. **Session-to-session learning.** Each run makes the next one faster and more effective. 3. **Identity spoofing.** The agents look like legitimate users at the session level. The attacks happen at the interaction layer, not the network layer, which makes them invisible to most current security stacks. Post-RSAC 2026, the industry's response focused on agent identity frameworks — verifying *who* an agent is. That's necessary but not sufficient. Knowing an agent's identity doesn't tell you what it's doing. Behavioral observation at the interaction layer is the gap that still needs filling. --- ## Where this leaves us The infrastructure for building agents is genuinely good now. Sandboxed compute, persistent state, cloud storage mounting, provider-agnostic SDKs — these are real capabilities that didn't exist in usable form 18 months ago. The harder problems are operational. Observability for reasoning chains, not just infrastructure. Governance that scales with agent count. Security models that account for what agents *do*, not just who they are. The first agent is easy. The fiftieth is where you find out whether you built the right foundations. ## Similar posts on daily.dev - [Tackling the Uncontrolled Growth of AI Agents in Modern SaaS Environments](https://daily.dev/posts/tackling-the-uncontrolled-growth-of-ai-agents-in-modern-saas-environments-xishdhvi7) · Security Boulevard · 0 upvotes · 0 comments - [Agentic AI Security: What CISOs Say About Governing AI Agents](https://daily.dev/posts/agentic-ai-security-what-cisos-say-about-governing-ai-agents-peda204z1) · Docker · 0 upvotes · 0 comments - [Google begins putting the guardrails on agentic AI](https://daily.dev/posts/google-begins-putting-the-guardrails-on-agentic-ai-gqgdwxx0s) · InfoWorld · 0 upvotes · 0 comments --- Tags: [#openai](https://daily.dev/tags/openai), [#ai-agents](https://daily.dev/tags/ai-agents) [View this post on daily.dev](https://daily.dev/posts/openai-agents-sdk-update-adds-sandboxed-workspaces-and-cloud-storage-support-hyrsyxgpz) ```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":"OpenAI Agents SDK update adds sandboxed workspaces and cloud storage support","url":"https://daily.dev/posts/openai-agents-sdk-update-adds-sandboxed-workspaces-and-cloud-storage-support-hyrsyxgpz","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/openai-agents-sdk-update-adds-sandboxed-workspaces-and-cloud-storage-support-hyrsyxgpz"},"datePublished":"2026-04-16T16:01:15.315Z","dateModified":"2026-04-22T15:35:36.802Z","description":"A survey of the current state of AI agent infrastructure, governance, and security threats heading into 2026. OpenAI's Agents SDK now supports sandboxed...","isAccessibleForFree":true,"articleSection":"Collections","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":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/openai-agents-sdk-update-adds-sandboxed-workspaces-and-cloud-storage-support-hyrsyxgpz","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":3},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"openai,ai-agents","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"OpenAI Agents SDK update adds sandboxed workspaces and cloud storage support"}]} ```