Browse by topic
AI Security
Security for AI agents and the infrastructure around them. Posts include a prompt injection delivered through a GitHub issue, a Cursor allowlist bypass, malicious agent skills, the backdoored LiteLLM release on PyPI, why an AI agent is often the most over-privileged account a team owns, and an AI agent alert that the logs did not support.
13 posts
-
6 min read
The fake browser extension playbook is back. This time it ships as agent skills
Malicious agent skills repeat the fake browser extension playbook: 341 of about 2,632 ClawHub skills were malicious. Five tells to check before you install a skill.
- AI Security
- Security Engineering
-
6 min read
My AI agent said we were under attack. The logs said otherwise.
My AI agent reported a prompt injection attack, but the logs showed none. How a forensic check exposed a false alarm and what I changed in my agent security.
- AI Security
- Security Operations
-
7 min read
No CVE needed: how a GitHub issue hijacked an AI agent
In a demo, one GitHub issue hijacked an AI agent via the GitHub MCP server and leaked private repo data. Why it needs no CVE, and what I changed.
- AI Security
- MCP Security
-
6 min read
The LiteLLM Supply Chain Backdoor: What 3 Hours Means for Your AI Agent Gateway
Two backdoored litellm versions, 1.82.7 and 1.82.8, sat on PyPI on March 24, 2026. What the window means for your AI agent gateway and how to pin builds.
- AI Security
- Security Engineering
-
3 min read
A fake MCP server spent three months earning trust. The tells were there
A fake MCP server from SmartLoader earned trust for three months. See which signals attackers can farm, which are costly to fake, and a pre-install checklist.
- MCP Security
- AI Security
-
6 min read
What Is MCP Security? Common Attacks and How to Scan Your MCP Servers
What is MCP security? Four attacks found testing 11 MCP servers: tool poisoning, command injection, path traversal, planted prompts, and how to scan for them.
- MCP Security
- AI Security
-
3 min read
The NSA published 17 pages on MCP security. Here it is as a checklist you can run today
The NSA's 17-page MCP security guidance as seven checks you can run today: inventory, pinning, tool descriptions, sessions, privileges, egress, logs.
- MCP Security
- AI Security
-
4 min read
A Happy-Path MCP Demo Proves Almost Nothing About Tenant Isolation
A happy-path MCP demo proves little about tenant isolation. What an MCP security preflight checks with static rules and behavioral tests, and what it cannot.
- MCP Security
- AI Security
-
9 min read
MCP 2026-07-28 Went Stateless: A Planted Prompt Is a Credential
MCP 2026-07-28 drops sessions and the initialize handshake, so a handle in the context window becomes a credential. Five per-request checks for server authors.
- MCP Security
- AI Security
-
5 min read
The Cursor Allowlist Bypass That Starts With a File Named curl
CVE-2026-22708 is a Cursor allowlist bypass that starts with a file named curl. I built a static shadowing check for it and explain two gaps it cannot close.
- AI Security
- Security Engineering
-
5 min read
The MCP attack your code review cannot see
MCP tool poisoning hides instructions in tool descriptions with invisible Unicode that code review misses. mcpscan is a static scanner with 12 checks.
- MCP Security
- AI Security
-
5 min read
Your AI agent is the most over-privileged account you own
An AI agent often gets full shell, your API keys and network on day one. Six least-privilege rules to contain prompt injection and shrink the blast radius.
- AI Security
- Identity
-
Medium ↗
I Built an AI-Powered SOC With Tools We Already Owned. Here's the Honest Version.
An honest write-up of building an AI-assisted SOC at Constellar with tools we already owned: what worked, what did not, and what I would do differently.
- Security Operations
- AI Security