Urgent.News

What's breaking now, across thousands of outlets.

Tech

readOnlyHint isn't a permission: blocking a destructive MCP tool by policy

MCP lets a tool describe itself. readOnlyHint: true says "I only read". destructiveHint: true says "I can delete or overwrite something". Clients can use those hints to decide when to ask a person before running a tool. Two things go wrong with that in practice. The hint is often missing. We looked at about 20,000 MCP tools for a report on MCP tool annotations . In Docker's MCP Catalog, 69% of…

MCP tools use hints to indicate whether they only read data or can delete or overwrite it. However, two issues arise with these hints in practice. First, they are often missing, and second, they can be incorrect. For example, a server in a catalogue marked 15 delete tools as read-only. The MCP specification advises clients to treat tool annotations as untrusted unless they come from trusted servers.

This suggests that the decision about what an agent may do should be made elsewhere, not in the tool's description of itself. A policy can be implemented that denies certain destructive actions regardless of the tool's annotations. For instance, a rule can be set up to deny the delete_customer tool, which permanently deletes a customer record.

This rule does not rely on the tool's annotations and simply denies the action. When the gateway receives a request to run the delete_customer tool, it applies the workspace's rules and responds with an error message, preventing the irreversible action. This approach ensures that irreversible actions require explicit approval, rather than being silently denied based on potentially incorrect or missing hints.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in Tech

Titanic Séance: ask the 891 passengers what the night would have done to you

Titanic Séance is a small web app where you tell the spirits who you would have been in April 1912: which ticket you held, how old you were, who you travelled with, and where you boarded.

  • Titanic Séance lets users experience life on the ill-fated ship
  • AI predicts survival chances based on passenger details
  • Gemma 4 language model voices real passengers

Website Security Best Practices Every Business Should Know in 2026

Most website breaches do not start with a clever zero-day exploit. They start with an outdated plugin, a reused admin password, or a form that trusts whatever a visitor types into it.

  • Outdated plugins, reused passwords, and unfiltered input are primary breach vectors
  • AI accelerates phishing and probing attacks on websites
  • Access control, updates, and monitoring remain core security principles

More from Tuesday 27 October →