Sourced / Security blog
AI & MCP Permissions
Website, API and AI security. Clear scope, useful evidence, and practical fixes.
30 articles / Page 1 of 2
- AI & MCP Permissions
Apply MCP Scope Step-Up
Request a stronger MCP scope only when a protected tool needs it, and deny the tool until new consent is verified.
- AI & MCP Permissions
Audit AI Tool Calls Without Secrets
Record agent tool decisions and downstream effects without storing raw credentials, private payloads, or transcripts.
- AI & MCP Permissions
Avoid MCP Token Passthrough
Keep client-to-MCP tokens out of upstream API calls by separating audiences, credentials, and per-user authorization.
- AI & MCP Permissions
Bound Multi-Agent Handoffs
Limit each child agent to a verified subset of the parent task's tools, data, time, and action scope.
- AI & MCP Permissions
Budget Agent Tool Operations
Set per-task limits for tool calls, external spend, data volume, retries, and writes, then test loop stopping.
- AI & MCP Permissions
Constrain MCP Sampling Requests
Review legacy MCP sampling requests for user approval, data limits, provider destination, and migration from deprecated sampling.
- AI & MCP Permissions
Contain Untrusted AI Tool Output
Keep retrieved pages, tool errors, and structured results from changing an agent's instructions or tool permissions.
- AI & MCP Permissions
Enforce RAG Retrieval Access Controls
Apply current user and tenant permissions before RAG chunks enter model context, including snippets and caches.
- AI & MCP Permissions
Gate High-Impact Agent Actions
Require a server-verified human decision before an agent can send, export, delete, refund, or change access.
- AI & MCP Permissions
Grade Agent Permission Tests
Score both permitted work and denied effects in synthetic agent tests, including incomplete and uncertain cases.
- AI & MCP Permissions
Grok Bot API Tool Safety Checklist
Limit Grok-connected MCP tools with an explicit inventory, caller checks, credential isolation, and safe test cases.
- AI & MCP Permissions
Keep Agent Secrets Out of Prompts
Keep provider credentials in a server-side broker and test that prompts, tool results, and logs never expose them.
- AI & MCP Permissions
Limit MCP Authorization SSRF
Constrain MCP OAuth discovery URLs, DNS results, redirects, and egress before the client fetches metadata.
- AI & MCP Permissions
MCP Servers: Architecture, OpenAPI Readiness, and Security
Understand MCP server roles and transports, assess OpenAPI tool readiness, and test permissions, credentials, and launch blockers.
- AI & MCP Permissions
Minimize MCP Client Scopes
Choose the smallest MCP scope set for first connection and test that protected tools stay unavailable.
- AI & MCP Permissions
Protect MCP Elicitation Inputs
Control the fields and URLs an MCP server may request from a user, with safe denial and phishing checks.
- AI & MCP Permissions
Record Per-Client MCP Consent
Bind MCP consent to the registered client, resource, user, and exact scopes; test changed-client and revocation cases.
- AI & MCP Permissions
Resolve MCP Tool Name Collisions
Find duplicate MCP tool names across servers and verify that names, handlers, and scopes remain distinct.
- AI & MCP Permissions
Review AI Tool Catalog Changes
Compare tool names, schemas, handlers, scopes, and results before a changed catalog reaches an agent.
- AI & MCP Permissions
Review Delegated Agent OAuth
Trace user consent through an MCP client, server, and upstream API with separate token audiences and object checks.
- AI & MCP Permissions
Screen RAG Ingestion Permissions
Record who can add RAG documents, how each source is classified, and which access rules survive indexing.
- AI & MCP Permissions
Separate AI Read and Write Tools
Split AI read and write tools by effect, with separate authorization, approval, and synthetic denial tests.
- AI & MCP Permissions
Stop Stale Agent Jobs
Recheck permission before queued agent work executes after a role change, cancellation, or lease expiry.
- AI & MCP Permissions
Test Prompt Injection in an Owned Lab
Build a synthetic prompt-injection case with a legitimate task, forbidden effect, positive control, and tool-event evidence.