Not a policy deck.
Every resource should give you a test, checklist, Skill, or workflow pattern you can use.
Agentic AI Security Field Guide
Secure the agents before they touch the workflow.
Practical research for humans and agents — MCP risk, prompt injection, identity, tool trust, and the controls that keep blast radius visible.
No pricing, SOWs, or service funnel. Just field notes you can use.
Not AI hype.
Workflow controls come before tool recommendations.
Not unmanaged automation.
Actions, approvals, and recovery paths stay visible before AI gets more authority.
Short answer
What does VibeSec Advisory help builders do?
VibeSec Advisory helps builders understand and test the risks created by agents, MCP servers, AI coding tools, prompt injection, and AI-generated workflows.
Read the latest field notesLatest field notes
Practical notes on agents, MCPs, prompt injection, AI coding tools, and safer workflows.
Essays, checklists, Skills, and field notes for people building and testing AI systems.
Agentic AI Security
Don’t Reuse Human Identity for Agents
Give the agent its own declared identity, bound to a principal chain that names who authorised it. Check declared versus observed behavior. Do not treat a reused human session, shared API key, or parent-token passthrough as the agent’s identity.
Read field noteAgentic AI Security
Bind MCP Approval to Content, Not the Tool Name
Bind approval to a canonical content hash of the executable configuration and tool definition—not to the tool name. If the hash changes, treat it as a new consent event and ask again.
Read field noteAgentic AI Security
Use the OWASP Agentic Top 10 as a Test Plan, Not a Badge
Use the OWASP Agentic Top 10 as a row-per-risk test plan with harmless probes, denial evidence, owners, and retest triggers before expanding agent access.
Read field noteThe problem
AI adoption without workflow design creates invisible risk.
People are using copilots, agents, and MCP-connected tools inside live business processes. The question is no longer whether AI is being used. The question is whether the workflow has a baseline, reusable skills, permissions, review points, and a measurement loop.
Workflow first
Start with the process, owner, inputs, approvals, and failure modes before tools.
Business metrics first
Good workflow review starts with the current metric, owner, and target state.
Guardrails by design
Human checkpoints, data boundaries, action limits, escalation, and security are mapped together.
Evidence before claims
Write down what you tested, what failed, what changed, and what still needs review.
Use The Field Guide
Start with the research, then use Skills when you need patterns.
The public front door is the research. Skills stay available as a real destination for reusable procedures and guardrails. Each resource should help you understand a risk, test an agent, review a tool, or improve a workflow.
Start here
Agentic AI Security Research
Read field notes on agents, MCPs, prompt injection, AI coding tools, generated code, and practical security workflows.
- Prompt injection tests
- MCP security notes
- AI red-team lessons
Operating artifacts
Skills
Browse reusable AI workflow skills, guardrails, examples, and implementation patterns that show how governed work should be captured.
- Reusable procedures
- Review gates
- Data boundaries
FORGE
A six-pillar model for agent-ready work.
FORGE remains a historical and educational model for thinking about governed workflows. The active VibeSec Advisory direction is a free field guide for securing agentic AI.
FORGE pillar
Baseline
Current workflow, owner, source system, pain point, and target metric.
FORGE pillar
Skills
Reusable instructions that capture how expert work should be done and reviewed.
FORGE pillar
Agents
Automation only where the tool boundary, action limit, and recovery path are known.
FORGE pillar
Guardrails
Data boundaries, approval gates, escalation rules, and security controls.
FORGE pillar
Schedule
Triggers, cadence, dependencies, and failure handling for repeatable work.
FORGE pillar
Capture
Evidence, lessons, regression checks, and the next measurable improvement.
How we work
Less strategy theater. More usable operating artifacts.
Capture
Capture reusable Skills, source boundaries, review gates, and test results so workflow knowledge compounds.
Map
Map agents to concrete steps, owners, tools, review points, and failure modes before trusting automation.
Contain
Guardrails are designed as blast-radius controls, not vague policy language.
Measure
Every artifact should include a way to test, review, or improve the workflow.
Start with the systems your agents can touch.
Read the latest field notes, then use Skills to test your own agentic AI systems.