The post introduces a Firecrawl kit for Docker Sandboxes that gives isolated coding agents live web search, scraping, and crawling through one allowlisted host, api.firecrawl.dev. It explains how the kit's mixin descriptor declares install, credential, phased network-policy, and agent-context capabilities, with the API key injected by a host-side proxy so it never enters the sandbox. It also covers verification steps and known limitations such as proxy 403s and organization policy overrides.
Docker engineering blog
Loading
The post contrasts containers, which share the host kernel, with Docker Sandboxes, which are microVMs with their own kernel and private Docker Engine built for autonomous coding agents. It explains the Docker socket exposure risk, how host-enforced network and credential controls work, and the setup and hardware costs of sandboxes.
The Score maintainers describe how the CNCF Sandbox project adopted Docker Scout, Docker Hardened Images, the Docker GitHub Builder, and OIDC connections through the Docker-Sponsored Open Source Program. The post walks through workflow and Dockerfile changes, CVE and package-count results, and how signed SLSA provenance and short-lived tokens replace long-lived credentials.
Docker and ActiveState announce a partnership that pairs Docker Hardened Images with ActiveState's Curated Catalog, a policy-governed repository of source-compiled open source packages. The post argues that governed package sourcing should accompany hardened base images to reduce supply chain risk from pip, npm, Maven, and AI coding agents. Most of the content is positioning and availability, with little implementation detail.
The post compares three ways to organize an AI agent's code in a Docker Sandbox: mounting a host checkout, giving each task a Git worktree, and keeping the repository inside a mountless sandbox. It covers trade-offs including shared dependency binaries across host and Linux sandbox, the limitation of mounting a linked worktree without its Git metadata, and the handoff workflow for bringing results back.
Docker describes its role as a founding member of the Blueprint Alliance and co-drafter of Agent Baseline, two reference efforts for securing AI agents. The post outlines new agent risks such as runtime reprogramming, borrowed identity, and containment that must revoke credentials, and frames the Alliance's four governance questions and Agent Baseline's six outcomes and 35 controls.
Docker's Part 6 of its AI Coding Agent Horror Stories series examines a read-only AI code review bot that fabricated a nonexistent CVE and recommended dangerous infrastructure actions. The post walks through moving the reviewer into a Docker Sandbox with a read-only code mount, no credentials, and a single allowed network path, plus automated NVD checks on CVE identifiers in agent output. It argues that agent output should be treated as untrusted input.
Docker announces its Navigator Partner Program, a three-tier partner structure (Crew, Explorer, Navigator) centered on AI governance co-selling, with benefits such as deal registration protection, MDF funding, and a temporary 10-point AI Governance Growth Incentive. The post frames Docker's role as the trusted runtime layer for running and securing AI agents, but it contains little technical detail about the platform itself.
Docker describes rebuilding its Docker Hardened Images catalog after two security events: the PHP2Shell flaws in the PHP engine and the eMudhra certificate compromise, which exposed intermediate CA private keys. The post explains why container trust stores keep trusting revoked certificates until images are rebuilt, and how SBOMs were used to verify CA store and PHP versions across published tags. The closing sections promote Docker's hardened image offerings.
Docker reports on two critical PHP memory-management vulnerabilities, CVE-2026-109618 and CVE-2026-109619 (PHP2Shell), which allow remote code execution in PHP core across supported and end-of-life versions. The post explains the failure modes: uninitialized memory read during cleanup after an out-of-memory abort, and a stale internal pointer reused in a later request. It lists patched PHP releases and the Docker Hardened and Official Images updates, and states that upgrading is the only...
Docker and ActiveState announced a reselling partnership that pairs Docker Hardened Images with ActiveState's curated open source packages to improve container security and dependency visibility. The post cites a survey and vendor claims about CVE reduction, remediation times, and SLAs, but offers little implementation detail. It reads mainly as a partnership announcement.
Docker and several security vendors published the Secure Software Factory reference architecture, which applies the Agent Baseline security outcomes to AI-driven software factories where agents write code, run tests, and hand results to reviewers. The central argument is that controls must live in the runtime environment around the agent, covering credentials, tools, dependencies, network access, and audit evidence, rather than relying on the agent to police itself. The paper also outlines...
Docker's blog introduces the official BAND AI kit for Docker Sandboxes, which lets an isolated agent join BAND collaboration rooms. The post explains the outbound-only WebSocket architecture, how credentials are injected by the sandbox proxy with placeholder values, and how per-sandbox network policies are scoped. It walks through a Repo Doctor example that repairs a broken Python project.
Docker's post cites 2026 Stack Overflow Developer Survey figures showing that only 7% of developers trust AI output for important decisions and that security and privacy rank as the second-biggest factor in AI tool adoption. It then promotes Docker Sandboxes and a $250 sign-up credit, offering little technical detail beyond the survey statistics.
The post outlines the main security risks of autonomous AI agents, including prompt injection, tool misuse and poisoning, excessive autonomy, credential exposure, and data exfiltration. It argues for containing what an agent can reach through isolation and scoped, short-lived credentials, and for governing its behavior with logging and action limits.
Docker introduces Harbor integration with Docker Sandboxes, running each agent evaluation task in its own microVM with its own kernel to isolate untrusted agent-written code. The post explains how configurable network policies can allow setup-time downloads and then cut off agent internet access, so evals stay fair and reproducible. It also covers local versus cloud execution, with concurrency tested at 64 workers.
Docker explains CVE-2026-105445, a Docker Hub account-conversion flaw that transferred active personal access tokens to newly named organization owners without consent or rotation. The advisory details the exposure window, investigation, remediation, token deactivation, and safer credential-management practices.
Docker introduces Cloud Sandboxes, which run AI agents in isolated microVMs and let developers move long-running work between local and cloud environments. The post also details the open Sandbox Kit specification, runtime-enforced permissions, auditability, and Docker’s plan to bring the format to CNCF governance.
Docker’s Sandbox Kit Specification v3 packages an AI agent’s workload, network policies, credentials, volumes, and lifecycle requirements into a pinnable OCI image. The post explains capability requests, deterministic mixin composition, conformance, and update-gating rules that make agent authority reviewable and reproducible across runtimes.
Docker introduces Cloud Sandboxes, extending its microVM-based environments for coding agents to Docker-managed compute. Developers can move sandboxes between laptops and the cloud with one command, run long-lived tasks in parallel, and apply isolated secrets, MCP integrations, and network policies.