Security Best Practices — MCP 2026-07-28
Versioned protocol guidance for authorization, token handling, SSRF and local-server risks. Start here when implementing an MCP service.
AN INDEPENDENT MAP OF AI SECURITY
Find your next
good source.●
Original research, practical guides, and trusted starting points. A smaller, more useful way into AI security.
— sources— selected readings08 topics
A question, a topic, or the source you already know.
Curated links. Original publishers.
EXPLORE THE INDEX
A little direction goes a long way.
The directory could not refresh. Your previous results are still here.
Model Context Protocol documentation
The protocol maintainers’ documentation is the starting point for MCP authorization requirements and transport security behavior.
MCP Security Cheat Sheet
A practical review checklist for MCP trust boundaries, authentication, tool access and deployment controls. Pair it with the protocol specification.
OWASP GenAI Security Project
Community security guidance, risk taxonomies and implementation resources. Use the stated edition when comparing framework entries.
The lethal trifecta for AI agents: private data, untrusted content, and external communication
A concise practitioner framework for recognizing agent data-exfiltration risk. Useful for deciding which capabilities should be separated.
Jumping the line: How MCP servers can attack you before you ever use them
Trail of Bits demonstrates how malicious tool descriptions can influence an agent before invocation. A useful threat-modeling case study.
Prompt injection is not SQL injection (it may be worse)
Explains why conventional injection analogies can mislead defenders, then discusses constraining agent capabilities and managing residual risk.
Trail of Bits AI and machine learning research
Original vulnerability research and audit engineering from a commercial security firm, with technical examples and project links.
Mitigating the risk of prompt injections in browser use
Anthropic describes browser-agent defenses and its own evaluations. Useful for understanding layered controls and remaining risks, with vendor-reported results.
MITRE ATLAS
An adversarial AI knowledge base of tactics, techniques and case studies for threat modeling and investigation.
Anthropic Research
Model-provider research on agent behavior, evaluations and defenses. Read reported results as vendor experiments with stated conditions.
Inside 90 days of attacks on AI infrastructure
Wiz reports observations from its AI-service honeypots, including exposed MCP services. Treat the telemetry as this vendor’s sampled observations.
Microsoft AI Red Team
First-party planning guides, lessons and tooling from Microsoft’s AI Red Team, useful for scoping and running assessments.
Planning red teaming for large language models (LLMs) and their applications
Microsoft’s planning guide covers assessment scope, people and process. Useful before selecting automated attack tooling or defining evaluation goals.
Not what you’ve signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
An early original paper demonstrating indirect prompt injection through retrieved content. Read the experimental setup and later revisions before generalizing.
Simon Willison’s Weblog
Independent practitioner explanations and experiments on prompt injection and agent design; distinguish demonstrations from general guarantees.
Wiz Research
Cloud-security vendor research covering AI infrastructure exposures and exploit chains. Separate disclosed observations from product marketing.
UK NCSC secure AI guidance
Government guidance for secure AI development across design, deployment and operation, with practical lifecycle considerations.
MITRE ATLAS tactics, techniques, and case studies data
The maintainers’ structured ATLAS data supports threat-model lookups and agent workflows. Preserve technique identifiers and linked case-study provenance.
GenAI and Agentic AI Exploit Roundup Q3 2026
An OWASP-hosted editorial roundup linking reported incidents and research. The page states it is neither exhaustive nor an official OWASP publication.
arXiv Cryptography and Security
Find recent security preprints and original papers. A posting is not peer-review approval; inspect methods, revisions and linked artifacts.
Building an AI Detection Engine That Understands Agent Intent
Analyzing model input and output logs in an AI-native detection pipeline to understand and uncover malicious AI agent behavior.
From Concept to Context Engine: How Wiz Built AI-Powered Data Discovery
Engineering article on the multi-agent pipeline and feedback loops Wiz used to turn a bucket scanner into an AI-powered data discovery context engine.
OWASP GenAI Security Project Unveils 2026 Top 10 for LLM Applications, New Agent Control Standard and Sponsors as Community Tops 30,000 Members
OWASP GenAI Security Project announces its 2026 Top 10 for LLM Applications, a new Agent Control Standard, and additional resources for securing generative and agentic AI.
Wiz Red Agent Finds Its Way Into Snowflake’s Internal Jira Through a Flaw in a GitHub Copilot–Assisted PR
Wiz Research article on an autonomous red-agent exercise involving a GitHub Actions injection in a Copilot-assisted pull request and an assessment of access to internal Jira.
A path we haven't mapped yet.
Try a broader topic or clear a filter.
“Newest” uses the original publication date. Undated guides remain available in curated order.
Directory revised
A USEFUL TOOL IN YOUR AGENT'S TOOLBOX
Let your agent start
with better references.
The same selected sources, with original links and clear dates.