← Back to the briefs
Tools

Anthropic adds background MCP tasks and strengthens agent permissions in Claude Code v2.1.319

Anthropic released Claude Code v2.1.319, adding terminal support for background MCP tasks, automated permission denial settings, and security fixes for worktree escapes and command bypasses.

Terminal MCP tasks and prompt automation

Anthropic updated Claude Code to version 2.1.319, introducing terminal support for Model Context Protocol background tasks. When an MCP server handles a tool invocation as a task, Claude Code shifts the execution to the background immediately and delivers the outcome to Claude when the job finishes. Operators who require synchronous execution can set an environment variable to force Claude Code to pause inside the tool call instead.

For unattended automated environments, the release adds a configuration that denies any tool invocation that would otherwise pause for user input. Once activated, the turn continues forward without interactive prompts across terminal runs, subagents, and collaborative teammate sessions.

Permission enforcement and auto mode adjustments

The update alters how auto mode behaves when telemetry or feature flags are disabled, including setups operating behind large language model gateways. If the remote server cannot validate a requested action, the system queries or denies the call outright, removing automatic fallback to the internal classifier.

Anthropic also modified the instructions provided to the agent when security boundaries halt a subagent. Claude is now advised to tell the operator or retry with a narrower prompt, rather than taking over the restricted task and executing the blocked steps itself.

Security patches for sandboxes and command evaluation

Version 2.1.319 delivers security fixes affecting execution sandboxes and rule evaluation. The update patches a sandbox escape vulnerability that occurred during worktree creation. Engineers also resolved a Bash permission bypass where crafted environment variable prefixes concealed command segments from permission checks, along with an issue that permitted stream editor commands to avoid checks during editing modes.

Our assessment

Teams running automated or continuous integration pipelines should test the automatic prompt denial configuration, as it keeps background tasks moving without human intervention. Operators relying on private gateways or disabled telemetry must verify their auto mode settings to ensure critical tool calls are not unexpectedly rejected under the stricter fallback policy.

Source: Anthropic original publication.

Original source

Read the original material

Anthropic (opens in a new tab)

The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.

Search the publication

Find a story by title, topic, or keyword.

Press Escape to close

Ask HorizonYour site guide
Your guide to the publication

Ask Horizon.

Ask a question about the publication.

Privacy