Agent Permissions and Least Privilege: The New Zero-Trust for AI Systems
Agents inherit user privileges in dangerous ways. The 2026 zero-trust patterns for agent permissions — scoped tokens, just-in-time access, audit.
Agentic AI, LLM engineering, and the models behind modern automation — multi-agent systems, LLM evaluation and comparisons, RAG, fine-tuning, AI infrastructure, security, and production AI engineering.
From the blog
Agents inherit user privileges in dangerous ways. The 2026 zero-trust patterns for agent permissions — scoped tokens, just-in-time access, audit.
Agent testing needs three layers — unit, integration, trajectory — and most teams ship only one. The 2026 test-suite blueprint that catches real regressions.
In agentic RAG the agent itself controls retrieval. The 2026 patterns and where they outperform classic retrieve-then-generate.
AI agents now participate at every SDLC stage. What changes in requirements, design, review, and deploy when agents are first-class collaborators.
Browser-using agents finally crossed a usability threshold in 2026. Anthropic Computer Use, OpenAI Operator, and the agentic web's emerging shape.
AIBOM is the SBOM equivalent for AI systems. The competing 2026 standards (CycloneDX-AI, SPDX-AI), what they capture, and which is winning.
Slack is the easiest place to deploy AI agents and the easiest place to get them wrong. The 2026 production patterns and pitfalls.
Incident reporting expectations changed in 2026. What OECD AIM, the AISI, and EU AI Office want from operators when an AI system fails.
Ticket routing, summarization, and resolution assistance in ITSM platforms. The 2026 patterns from real ServiceNow and Jira deployments.