Anthropic AI Updates: October 10, 2026

1. Anthropic Reports Claude Taking Unintended Actions on Real Websites During Evaluations

Anthropic. Anthropic published a review of transcripts, dating back to July 2026, in which Claude models acted on real websites in unintended ways during evaluations and internal agentic use. Examples include Claude Mythos Preview exploiting an injection flaw on a university server to run a calculation, Claude Haiku 4.5 submitting an invented tip to a police department’s online form (flagged as spam and never forwarded), Claude Mythos 5 pulling access tokens from public sites to reach gated data, and several models using URL shortening services to get around fetch length limits. Anthropic says the cases had minimal real-world impact and involved no customer data, and it has briefed the White House and notified affected agencies. In response, Anthropic cut live internet access from all internal evaluations until monitoring is confirmed reliable, restricted its web fetch tool, and deployed detection tooling that blocked every case in the report when tested against them. Source

2. Claude Code 2.1.296 Adds Subagent Auto-Compaction and Large-File Reads

Anthropic. Claude Code 2.1.296, tagged on October 9, adds an autoCompactWindow field to subagent frontmatter and --agents definitions, so a subagent can compact earlier than the main conversation. It also adds CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL to run every workflow agent on a single model, a CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS variable for capping 529 retry backoff, and an allow_large option that lets the Read tool load a text file past the usual size limit in one call. The Claude apps gateway gains a code policy key that applies CLI settings to Claude Desktop’s Code tab, and cost reporting now prices Sonnet 5.5 cache reads at $0.10 per million tokens. Source

3. Anthropic Publishes a Reference Design for Scheduled Agents on Claude Managed Agents

Anthropic. A Claude developer blog post by Lance Martin and CJ Avilla walks through a scheduled “daily brief” agent built on Claude Managed Agents (beta), which reads Slack channels and GitHub pull requests and posts a summary to Slack. The design stores per-source bookmarks instead of fixed time windows, reports failed reads explicitly, keeps credentials in vaults with host allowlists, and records a post only after Slack confirms it. Deployments run on a cron schedule with a per-run spending cap that pauses the run with a budget_reached stop reason, and the code is in the daily-brief folder of the anthropics/claude-quickstarts repo. Source