All topics

AI agent security: agents, MCP, and prompt injection

AI agent security starts with the tools, credentials, and permissions that let a model act. This guide is for AppSec engineers and security leaders deciding how much authority to give agents and how to enforce its limits. AI Security Table, the roundtable podcast with Chris Romeo, Izar Tarandach, and Matt Coles (formerly The Security Table), debates agent access, approval fatigue, MCP, and prompt injection through the conversations below.

Questions we keep arguing about

How much access should an AI agent have?

The hosts apply least privilege to agents that can request permissions and take actions. They question whether users can judge a request for administrator access, especially when repeated prompts make approval routine. The access boundary needs to limit what the agent can do, rather than depend on its explanation. Hear the discussion.

Is human approval a real control?

The hosts debate how approval fatigue turns human involvement into a false sense of security. An approval step loses value when people accept changes without understanding their consequences. Their discussion asks what the person can actually inspect before agreeing. Hear the discussion.

What can go wrong with MCP?

The MCP conversation examines the connections between models, tools, and other agents. The hosts question how these connections change the security assumptions around a system. Connecting another tool means considering another route through which the model can act. Hear the discussion.

How does prompt injection reach an agent?

Sander Schulhoff explains how crafted input can push a model toward unintended outputs. The hosts use HackAPrompt examples to examine why a model can resist one phrase yet remain open to other manipulations. The discussion connects those input risks to the need for limits on AI decision making. Hear the discussion.

Episodes

The Agentic Access Problem: When AI Becomes Its Own Administrator

Jun 3, 2026 · 40 min

The hosts explore agents that request permissions and the uncertainty behind an administrator-access prompt. They connect least privilege to prompt fatigue and responsibility for automated actions.

Read the transcript

The Tool Creep Problem: When More Security Means Less Security

May 8, 2026 · 42 min

Izar describes an agent attempting workarounds around a sandbox. The table argues for defined boundaries and limited capabilities while questioning whether adding more tools improves security.

Read the transcript

The Human In The Loop Illusion: Why AI Approvals Are Failing Security

Apr 30, 2026 · 48 min

A debate about human approval examines the gap between having a person in the loop and having meaningful oversight. The hosts connect approval fatigue to the review of AI-generated code.

Read the transcript

The Agent Access Problem: When AI Has the Keys, Who’s Really in Control?

Mar 25, 2026 · 48 min

Credentials, APIs, and permissions make agents actors inside existing systems. The table questions whether new governance lists address over-permissioned agents and unpredictable behavior.

Read the transcript

MCP…Something Could Go Wrong

Jun 3, 2025 · 46 min

The hosts unpack the Model Context Protocol and the security questions raised by connecting models to tools. They extend the discussion to communication between agents and the limits of AI safety.

Read the transcript

The Impact of Prompt Injection and HackAPrompt_AI in the Age of Security

Dec 19, 2023 · 65 min

Sander Schulhoff introduces prompt injection through the HackAPrompt competition. The conversation explains how manipulated input changes model output and why AI decision making needs security controls.

Read the transcript

Related

All conversations in this topic

14 conversations · newest first

Explore the full episode archive