AI security · Updated
AI security for apps, agents, MCP servers and AI-written code
AI changes security in two ways. Apps and agents built on language models can be steered by text an attacker writes, and code written with AI assistants arrives faster than teams can review it. This hub collects our guides on both, each built on primary sources, and the reviews and audits we run to find and fix the problems.
By George Shvaya, founder of Webvello · Published
Key takeaways
- Prompt injection is first on the OWASP Top 10 for LLM Applications, and OWASP, the UK NCSC and the model vendors all say it cannot yet be fully prevented.
- The MCP specification documents risks including confused deputy, token passthrough and SSRF, and two MCP developer tools had critical CVEs in 2025.
- Veracode’s 2025 tests found AI-generated code introduced security vulnerabilities in 45% of cases.
- CISA’s KEV catalog listed 1,781 vulnerabilities exploited in the wild as of November 16, 2026, the first signal for which dependency alerts to fix.
- The defenses are mostly ordinary security engineering applied consistently: least privilege, review, validation and patching.
AI security services
- AI application security reviewPrompt injection testing, tool permissions, data exposure and output handling for AI apps and agents, mapped to the OWASP LLM Top 10.See the service
- MCP server security auditAuthorization, token handling, tool scopes and request validation, checked against the MCP specification’s security requirements.See the service
- AI code and dependency auditInsecure patterns, hallucinated and malicious packages, leaked secrets, triaged dependency alerts and coding agent settings.See the service
AI security guides
- What is prompt injection?Direct vs indirect injection, why it cannot be fully prevented, and how to limit the damage.Read the guide
- MCP security risksTool poisoning, confused deputy, token passthrough, SSRF and two critical 2025 CVEs.Read the guide
- OWASP Top 10 for LLM applicationsAll ten 2025 risks in plain English, with a first defense for each.Read the guide
- AI coding assistant securityWhat studies found about AI-written code, hallucinated packages and leaked secrets.Read the guide
- Dependency vulnerability triageKEV, EPSS, reachability, the NVD backlog and the 2025 npm attacks.Read the guide
- Bank DNS security studyRegistry lock, DNSSEC, CAA and DMARC on the domains of the 100 largest U.S. banks.Read the guide
- DNS attack or outage?The public evidence a DNS attack leaves, and what the November 16, 2026 bank outages showed.Read the guide
Not sure where your risk is?
Tell us what you are building with AI. We will tell you which of these risks apply and what to check first.
Which AI security risks matter most right now?
| Risk | Where it shows up | Guide |
|---|---|---|
| Prompt injection | Chatbots, RAG features, agents reading email, web pages or files | Prompt injection |
| Excessive agency | Agents with broad tools, credentials or no approval steps | OWASP LLM Top 10 |
| Tool poisoning and token passthrough | MCP servers and the assistants that connect to them | MCP security risks |
| Insecure generated code | Code written with Copilot, Claude Code, Cursor or Codex | AI coding assistant security |
| Hallucinated and malicious packages | Dependencies suggested by AI or updated automatically | Dependency triage |
How we approach AI security
- We start from what the model can reach, because that sets the worst case for every other risk.
- We follow public standards: the OWASP Top 10 for LLM Applications, the MCP specification, and CISA and FIRST guidance for vulnerabilities.
- Findings come with evidence and a fix, ranked by real risk rather than scanner severity alone.
- We test only systems you own or are authorized to test, and we prefer staging environments.
Webvello’s founder, George Shvaya, studied cybersecurity and information assurance and comes from networking and information security. We also build the systems we review: MCP servers, AI agents and Claude Code rollouts. Rules that affect AI tools, such as the FCC’s ruling on AI voices, are in our rules tracker; the OWASP GenAI Security Project is the best place to follow the field.
Related services and resources
Pages that go deeper on the pieces this one touches.
Common questions
Straight answers to what owners ask us first.
What is AI security?
AI security covers the risks that come from building with AI models: attackers steering a model with injected text, agents with too much access, data leaking through prompts or retrieval, and insecure code or packages introduced by AI coding tools. It sits alongside ordinary application security rather than replacing it.
Do small companies need an AI security review?
If an AI feature can read customer data, send messages, change records or run code, yes. A review scoped to one feature is small, and it is cheaper than finding out from a customer.
What happens after I ask for a free security review?
You describe your app, agent, MCP server or codebase. A person replies with what we would check first and what a full review or audit would involve.
Find the gaps before someone else does
Book a call or describe your app, agent, MCP server or codebase, and we will scope a review.