Obsidian AI Agent Security Academy

The practitioner's guide to securing AI agents

From MCP server risks to machine identity and runtime guardrails — everything security engineers and CISOs need to govern agentic AI without slowing adoption.

xxx
Articles
·
xxx
Security Pillars
·
Updated
June 1, 2026
All (0)
Visibility & Shadow AI
Definition
ToF
Mcp Security

What Is MCP Security? A Security Team's Guide

The Model Context Protocol (MCP) is Anthropic's open standard for connecting AI agents to external tools, data sources, and services. Thousands of MCP servers are now running inside enterprise environments, and the Vulnerable MCP Project tracks more than 50 known vulnerabilities across them, 13 rated critical. The hard truth: most of those servers are completely invisible to traditional security tooling.
Access & Permissions
Threat Explainer
MoF
Ai Agent Toxic Combination

AI Agent Toxic Combinations: When Risk Stacks Reach Critical Severity

Most AI agents running in enterprise SaaS environments hold more permissions than their workflows require. That single finding rates as medium severity. Add an orphaned creator account and a public-facing deployment, and you no longer have a medium-severity finding. You have a critical incident waiting for a trigger. That is the core problem with how most teams score <a href="https://www.obsidiansecurity.com/blog/ai-agent-security-risks">AI agent risk</a> today: they evaluate each factor in isolation and miss the compound severity that combinations create.
Access & Permissions
Threat Explainer
MoF
Ai Agent Overpermissioned

Overpermissioned AI Agents: The Excessive Access Risk Security Teams Overlook

Most AI agents running in enterprise SaaS environments today hold more access than they need to do their job. That finding alone is striking. But here is the hard truth that makes it structurally different from human over-permissioning: when a human employee accumulates too much access, a quarterly access review can catch it. When an AI agent does, it is already moving data at machine speed across dozens of connected systems, with no manager to flag unusual behavior and no insider risk program watching its actions.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
300×
increase in AI agent activity, 2025
40%
of agents carry medium-to-critical risk
800+
risky agents in avg enterprise environment
25–50×
native tools built to see agent identity
Source: Obsidian Security Research, 2026

The six pillars of AI agent security

Get new AI agent security research in your inbox.
New articles, threat research, and CISO briefings — no spam, unsubscribe any time.
See every AI agent in your environment — in 30 minutes.
Obsidian surfaces agent inventory, toxic combinations, and MCP server activity across 8 platforms. No connector required.
Get a Demo
No setup required · 30-min session

Trusted by leading security teams

AI agent security: frequently asked questions

What is AI agent security?

AI agent security is the practice of governing, monitoring, and enforcing policy over autonomous AI systems — agents — that take actions inside business applications without direct human instruction for each step. It differs from traditional application security because agents hold persistent credentials, execute multi-step workflows, and access sensitive SaaS systems at a scale and speed that human oversight cannot track manually.

How is AI agent security different from prompt security?

Prompt security governs what human users submit to language models: filtering sensitive data, detecting injection attempts, blocking harmful outputs. AI agent security governs what autonomous agents do after they receive their instructions — which credentials they use, which SaaS applications they access, which actions they chain.

What is a toxic combination in AI agent security?

A toxic combination occurs when multiple medium-severity risk factors exist simultaneously on a single agent, compounding into a critical-priority finding. Example: embedded admin credentials + org-wide public access + disabled creator account = critical blast radius, no owner.