Security and sandbox policy fixes
Anthropic patched several security boundaries in Claude Code v2.1.321. The tool previously exposed sensitive authentication material when an MCP server issued an HTTP redirect, allowing OAuth credentials like refresh tokens to escape to separate hosts. Anthropic also addressed sandbox handling for shell execution, correcting a failure that allowed read-only Bash commands to bypass the sandbox without prompt verification.
File manipulation controls received updates to prevent unauthorized writes. Bash permission checks previously misjudged destination folders when validating write requests, omitting needed user prompts. In addition, directory level copy and move operations now trigger file protection rules and deny checks on destination files, while hook outputs exceeding 16 MiB now cause tool operations to fail safely instead of skipping evaluation.
Remote session administration and model defaults
The release introduces Claude Fable 5.2 as the default model for Fable configurations, supporting a 1M token context window. Anthropic also expanded administrative flags for remote sessions. Operators can now supply a dedicated configuration file using the settings parameter for remote control tasks. If the configuration changes on disk, new sessions halt until the process restarts.
Subagent orchestration was aligned across runtime environments. In headless execution, VS Code, and desktop sessions, agent reports waiting concurrently from multiple background subagents now batch together into a single model query rather than triggering split round-trips.
Our assessment
Teams relying on Claude Code for automated coding workflows should review active MCP server configurations, especially when connecting to external authentication providers that issue cross-host redirects. Administrators deploying remote control instances should adopt the settings flag to lock down uniform runtime behavior. Because copy and move actions now strictly enforce path protection rules, developers should verify that internal build scripts or plugin hooks do not unexpectedly stall on newly restricted target directories.
Source: Anthropic original publication.
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.





