Agent Security: articles and guides

AI agent security tools and platforms compared: Zenity, Noma and PlayCISO labs for permissions, approvals and containment. Read 4 Nov 2026.

Anthropic’s open-source sandbox-runtime (srt) confines an AI agent’s filesystem and network without a VM. How it works, its limits, a 13-point checklist.

India’s CERT-In (CIAD-2026-0049) warned AI agents under test reached live organisational systems. What it said, and the least-privilege fixes to apply.

Muxin runs AI coding agents across all your machines from one window. Useful, but it concentrates shell access in one app. The risks and safe use.

swarmcha.se reports AI agents on Tencent Cloud scraping Amap data, 28 Sept-8 Oct 2026. What was reported, the authors’ caveats, and the defender lessons.

A practical guide to designing, securing, and optimizing MCP servers — covering tool-poisoning defenses, least privilege, and what to deploy.

MCP servers give AI agents the same trust as internal tools — here's how to secure them against tool poisoning and prompt injection.

Google Labs’ Project CC gives an AI assistant its own account and Gmail, Calendar and Drive access. The agent identity, approval and injection lessons.

Open Dot is the open-source OpenAI Dots alternative: always-on AI agents on your Mac (composio-community/open-dot). How it works, cost and the risks.

When you give an LLM the ability to call tools and take actions, you grant it real capability. The controls that keep LLM tool access safe: least privilege, human approval, sandboxing and logging.

The Model Context Protocol lets AI agents call external tools and data through MCP servers. The governance controls that prevent an over-privileged or malicious MCP server becoming your next incident.

MCP tool poisoning hides malicious instructions in tool descriptions to hijack AI agents — here's how the attack works and how to defend against it.

Why individually correct LLM-agent security controls still leak when composed — the "security-context discontinuity" failure mode across tool calls and agent boundaries, and how carrying authenticated context end-to-end fixes it. A plain-English CISO read of the CONTINUITY research (zero harmful effects across 2,560 attacks).

A category-honest comparison and ranking of AI-agent command sandboxes in 2026 — Anthropic's srt against bubblewrap, Firejail, nsjail, Landlock, Seatbelt, Deno/Bun, Docker, gVisor, Firecracker, E2B, Daytona and others.