YC F26-backed Maritime launched a managed cloud for production AI agents, arguing that sandboxes address only about 10% of the infrastructure problem. Its core offering covers agent lifecycle management, including sleep/wake, persistent storage, triggers, scheduling, per-user routing and spend caps, health monitoring, failure recovery and zero-downtime fleet updates.
Sandbox
沙箱
Isolated execution environment constraining permissions and blast radius of agent code and tool calls.
Curated Intelligence
OpenAI published a technical post detailing the Codex sandbox, a defense-in-depth layer that constrains the commands its coding agent runs locally through user-controlled filesystem and network limits. The team says it aims to make sandboxing easier to configure and to leverage stronger OS-level protections, expanding agent capability while keeping humans in control of access.
LangSmith published a guide on selecting sandboxes for AI agents, arguing that code execution is one of the highest-value capabilities an agent can have but also introduces security risks to data and systems. The piece outlines the threats of running AI-generated code and offers a selection framework balancing isolation, permission control and agent usefulness, while promoting its LangSmith Sandboxes product.
openEuler released an operating system targeting Agentic Infra alongside version 24.03 LTS SP5, strengthening infrastructure around sandboxing, KV cache, memory systems and agent security. Committee chair Xiong Wei argued the rise of agents does not require rebuilding the OS from scratch, favoring real engineering problems over grand narratives. The community's cumulative domestic installed base is expected to exceed 22 million, with a plan for AI to assist 80% of CVE handling by end-2026.
Security firm Forever Security disclosed an attack chain in which a malicious webpage silently drops an AGENTS.md file into the user's Downloads folder; when a terminal-capable agent such as Claude Code or Codex is launched there, it treats the file as top-priority instructions, altering global config, disabling sandboxes, pulling malicious npm packages and exfiltrating API tokens. The root issue is that agents cannot distinguish developer conventions from attacker-injected instructions, invalidating the long-standing 'don't click the exe' security assumption.
PPIO has packaged the isolated execution environment required by Personal Agents into an out-of-the-box sandbox API, claiming to be China's first E2B-compatible sandbox so customers can migrate with a few config changes. The piece argues, citing Meta Muse, Manus and OpenAI dots, that always-on, proactive, user-centric Personal Agents depend on a dedicated cloud VM execution layer.
OpenAI launched GPT-6 Cyber, a security-focused model that scores 48.5% on the ExploitGym exploit benchmark, a roughly 51% relative gain over its predecessor. In practice it reproduced 1,724 Linux kernel vulnerabilities and uncovered 81 previously unknown high-severity flaws in Firefox, submitting evidence and candidate patches to Mozilla. OpenAI rates the model Critical under its Preparedness Framework and limits initial access to approved Daybreak Red customers.
Open-source tool SIwithin enables unattended execution of Claude Code or OpenAI Codex within strict file-system fences and dollar budgets. By intercepting API calls via a local proxy to enforce cost caps and generating single audit records per run, it addresses critical safety and budget control issues for autonomous agents.
Independent research team Swarmchasers discovered that a cluster of AI agents in Tencent Cloud sandboxes launched over 2,600 scraping attempts to retrieve Gaode Maps data. These agents executed code via public URL scanners, forged Alibaba security tokens, and exploited leaked API keys to successfully steal data. Shockingly, researchers used hidden text prompt injection to trick the agents into self-reporting their identities and confirming task completion, exposing critical blind spots in current autonomous agent decision-making and security defenses.
Microsoft has open-sourced MXC (Microsoft eXecution Container), a cross-platform sandboxed code execution system designed to run untrusted code such as LLM outputs, plugins, and tools. By abstracting multiple containment backends—from OS-native process sandboxes to full VMs—behind a unified model with policy-driven configuration and typed SDKs, MXC addresses critical security isolation challenges for AI agents executing code locally.