Securing agent tools and MCP servers
Spot confused-deputy and tool-poisoning risks in an agent's tools and MCP servers, and add allow-lists, scoped credentials, approval screens and logging so a bad tool or tool result cannot cause real harm.
Pro 1 · Quality engineerabout 3 hours5 sections
Quick answer
Tools give agents real power. A poisoned tool description, a hostile tool result or an over-broad token can turn the agent against its own user, so vet, pin, scope and gate every tool.
Why it matters
Agents now connect to email, code, payments and company data through tools and MCP servers, many written by strangers. One malicious server or one tool result with hidden instructions can make a trusted agent leak data or act for an attacker. Security teams review exactly this.
Sign up free to read the whole lesson
The rest of this lesson: 5 parts, video lessons, the best course, practice and a quick check. It is free, takes a few seconds and needs no card.
Accounts are for learners aged 18 and over. Under 18? The quick answer and the first part of every lesson stay open: browse lessons by topic.
Lessons on AI Security Engineer0 of 40 done
- 1What AI, machine learning and a language model are
- 2How cyber attacks work
- 3Networking for security
- 4Linux for security
- 5Variables, loops and functions in Python
- 6The terminal and Git
- 7Web app security and the OWASP Top 10
- 8Identity and access
- 9Cryptography basics
- 10Logs and SIEM
- 11Threat detection with MITRE ATT&CK
- 12Incident response basics
- 13Phishing and social engineering
- 14Cloud security basics
- 15Audit logs
- 16How AI products work
- 17Machine learning for threat detection
- 18LLMs in the SOC
- 19AI-powered attacks and how defenders respond
- 20Deepfakes and misinformation
- 21Threats to LLM apps: the OWASP Top 10
- 22Defend against prompt injection
- 23Guardrails
- 24Data quality and personal data
- 25Permission-aware RAG
- 26Security and privacy
- 27Data poisoning, backdoors and adversarial examples
- 28Model extraction and privacy attacks
- 29Securing the AI supply chain
- 30Securing agent tools and MCP servers
- 31Permissions and sandboxing
- 32Threat model an AI system
- 33Red-team an AI system
- 34Test sets and evals
- 35Security tests in CI for AI apps
- 36Run an AI security programme
- 37AI risk frameworks
- 38AI rules in India and abroad
- 39Incident response for AI
- 40Security questionnaires
Related
On the roadmap: AI Security Engineer
SecurityPro 1: Quality engineerVideo lessonsTool poisoningConfused deputyModel Context ProtocolExcessive agencyLeast privilegeHuman in the loop