FTC Demands Documents From AI Giants
Claude Code now compacts long idle sessions before prompt cache expires, and headless sessions default to standard permission mode instead of auto. Anthropic's Q3 revenue hit $16.8 billion ahead of a planned IPO, the FTC subpoenaed the company over safety incidents, and Replit ended its partnership with Anthropic.

AI in your oven? Anthropic warns models ‘months away’ from self-replication on smart devices — The Guardian
Highlight
Idle compaction keeps long sessions from reprocessing
If you leave a 200K-plus token conversation sitting for about 54 minutes, Claude Code now compacts it just before the one-hour prompt cache expires, so your next message doesn't pay to reprocess the whole history. You can tune the threshold with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, turn it off with "idleCompaction": false in settings.json, or toggle it from the new Idle compaction row in /config.
Worth Knowing
Headless sessions now start in default permission mode
If you run claude -p or the Agent SDK without specifying a permission mode, sessions on third-party providers, with telemetry off, or on a fresh install now start in default rather than auto. Pass --permission-mode auto if you were relying on the previous behavior.
MCP permission rules now match only the named server
An allow rule like mcp__a no longer leaks across to servers named a__b or a_ — write mcp__<server>__* with the server's full normalized name instead. The same applies to an agent's tools: list, so check any MCP rules with underscores in the server name.
Fork subagents become the default outside the terminal
Fork subagents are now on by default in claude -p, the Agent SDK, the desktop app, and VS Code, matching the terminal, and the desktop app and VS Code now always run your main conversation's subagents in the background instead of letting Claude choose. Set CLAUDE_CODE_FORK_SUBAGENT=0 if you want the old behavior.
You can stop file mentions from inlining contents
Set CLAUDE_CODE_DISABLE_FILE_MENTIONS and an @ mention hands Claude the path instead of dumping the file into your prompt, letting it read selectively with its file tools. Useful if you mention large files and want to keep context spend down.
VS Code renders LaTeX, and vim mode gains bracket jumps
Claude's responses in VS Code now render LaTeX math, including $$ blocks and inline $…$, and the new claudeCode.attachSelection setting stops editor selections from being appended to your messages. In the terminal, vim mode picks up the % motion for jumping to a matching bracket, on its own or with d/c/y.
OpenTelemetry traces gain GenAI semantic conventions
Trace spans now carry standard attributes like gen_ai.provider.name, gen_ai.conversation.id and error.type, tool_decision events add classifier_verdict and permission_mode, and api_request splits cache writes into cache_creation_5m_tokens and cache_creation_1h_tokens. If you are wiring this into a backend, the new OpenTelemetry tracing cookbook walks through the spans attribute by attribute.
Under the Hood
Sandbox, plan mode and credential leaks closed
Plan mode no longer runs writes that a PreToolUse or tool.check hook allowed, sandboxed commands on macOS can no longer reach a denied path through a symlink or write into a --plugin-dir folder, and feature flag, startup and fast mode requests no longer carry a third-party gateway's API key to Anthropic. The Read tool also now refuses image sources over 128 MB instead of loading them whole into memory.
Startup, streaming and memory improvements
Plugin command and agent files load with far fewer filesystem calls, which speeds startup on Windows in particular, the VS Code panel stops re-parsing the full thinking text on every update, and unbounded memory growth under --debug logging and after resuming subagent-heavy sessions is fixed.
From Anthropic
In the News
The permission default changes for headless sessions and MCP allow rules are the ones most likely to surprise you — worth a look at your existing rules if you script Claude Code.




