531 posts tagged with "AI agents"
AI agents and autonomous systems
10,000 MCP Servers Later, the Hard Part Isn't Building One — It's Being Found and Trusted
MCP passed 10,000 public servers and 97 million monthly downloads in 2026 — but 52% of those servers are abandoned. Here's what building bex's own MCP server into that ecosystem actually requires: the protocol's stateless rewrite, the tool-poisoning threat class, and why shipping one isn't a differentiator anymore.
Backstage Ships MCP Tokens for Claude and Cursor: What Exposing a Golden-Path IDP to AI Agents Actually Requires
Backstage 1.43 lets Claude and Cursor call your golden-path Scaffolder actions with a short-lived, per-user token instead of a shared secret. Here's the actual config, the failure mode an unscoped action list creates, and what a self-hosted PaaS's own MCP server needs to copy.
Better-PaaS Ships Scoped Agent Tokens for Cursor and Claude Code — What an AGPL Self-Hosted PaaS Choosing Managed Databases Over bex's Non-Goal Reveals
Better-PaaS, a single-maintainer AGPL project, ships three-tier scoped agent tokens and seven MCP tools for Cursor and Claude Code — a concrete look at what that buys against bex's own coarser key model, and where Better-PaaS's bundled managed databases trade convenience for the high-availability story a single VPS can't back.
Three JavaScript Quirks, One CVSS 10.0 RCE: What n8n's Sandbox-Escape Chain Means for Every Agent Tool Wired to Your Cluster
Three individually-harmless gaps in n8n's JavaScript sandbox chained into a CVSS 10.0 RCE that reached every stored credential and, on shared instances, the Kubernetes cluster underneath. Here's the exploit chain and what it means for any tool that hands an agent a general-purpose sandbox.
The NSA Just Published a Threat Model for MCP: What It Means for Your Deploy Agent
The NSA's AI Security Center named four structural risks in MCP's design — and every one already has a real 2026 incident behind it. Here's what each means concretely for a self-hosted MCP server with deploy and rollback authority.
Gitpod Bet the Company on Self-Hosted Agent Infra — Then OpenAI Bought It
Gitpod rebranded to Ona and went self-hosted-only in September 2025. Nine months later OpenAI bought it for Codex. Here's what a hyperscaler choosing to buy customer-owned execution instead of building it actually signals.
What a Zanzibar-Style Relationship Graph Buys a Deploy MCP Server Over a Scoped API Key
AuthZed is pitching SpiceDB's relationship-graph authorization for AI agents. Here's the actual SpiceDB schema for a deploy/rollback MCP server, what it can check that a scoped API key can't, and whether running it is worth the operational cost.
SPIRE Needs Your Fleet's Shape in Advance. An Orchestrator Spawning Sub-Agents Doesn't Have One.
SPIRE requires every workload to be pre-registered before it can attest — a model that breaks the moment an orchestrator spawns sub-agents on demand. Here's the actual registration race, the ClusterSPIFFEID fix, and what it still doesn't solve.
Your Deploy Agent Has the Same Privileges as a Human Push — and No Insurance
AIUC-1 and California's AB 316 just made 2026 the year AI agent liability got real, while insurers quietly excluded it from standard coverage. Here's what's actually covered, what isn't, and the audit-trail and permission design a deploy-from-chat platform needs today.