Skip to main content
Back to all releases
// November 13, 2026v1.52.4

Platform

Security and Policy

Risk policies now catch tool calls before they run, in Claude and in Claude Cowork

Tool calls used to slip through policy enforcement in two places: Anthropic inference hooks only scanned prompts, and Claude Cowork didn't scan anything in real time at all, so a block or warn only showed up later as a finding, after the call had already run. Both gaps are closed. Platform MCP chat search also masks more kinds of sensitive data now. More on this in Guardrails.

Features

  • Risk policies now scan tool calls through Anthropic inference hooks #7788 - Each tool call Claude produces is now scanned with the same policies as prompts, and a block, warn, or quarantine match denies it before any of its tools run. Previously these calls went through unscanned, so a policy violation was only caught on the next prompt, after the call had already executed. (Author: @speakeasyforgebot)
  • Risk policies now enforce in real time in Claude Cowork #8932 - Prompts and tool calls in Cowork are now scanned as they happen, the same as in Claude Code. A warned prompt gets a short approval link, and a warned tool call gets Claude's own permission prompt, instead of only showing up as a finding after the fact. (Author: @mfbx9da4)
  • Chat search masks more kinds of sensitive data #8386 - Platform MCP chat search now masks emails, phone numbers, card numbers, SSNs, medical record numbers, URL passwords, and one-time codes in transcripts, not just the credentials the secret scanner finds. (Author: @simplesagar)
Sagar Batchu
Sagar Batchu
View on GitHub