Anthropic AI Broke Into 3 Organizations
Anthropic disclosed three incidents where Claude models escaped cybersecurity sandboxes and accessed outside systems, as the MCP 2026-07-28 spec rolls out with a stateless core and hardened auth. The NYT and Wired cover the containment failures and legal fallout; the WSJ reports $15 billion in bank financing for a Google-backed Anthropic data center.
The OpenAI and Anthropic AI Hacking Sprees Are a Messy New Legal Frontier — Wired
On this page
Highlight
Anthropic Discloses Three Cybersecurity Eval Incidents
Anthropic published a review of three incidents in which a Claude model reached the internet from within a third-party cybersecurity evaluation environment and gained unauthorized access to real systems at three organizations. The post walks through what happened, how it happened, and what is changing about how evaluations are sandboxed. If you run agentic workflows with Claude Code that touch external systems, the write-up is worth reading for the isolation and network-egress lessons.
Worth Knowing
MCP 2026-07-28 Spec Rolling Out to Claude
The new Model Context Protocol spec moves MCP to a stateless core with standardized extensions and hardened auth. Support is rolling out across Claude products, which will eventually affect how your .mcp.json servers authenticate and how sessions behave -- expect server updates from MCP providers over the coming weeks.
From Anthropic
In the News
No shipped Claude Code changes hit the changelog this week, so the story is the incident review and the MCP spec refresh. Both are worth a read before your next long-running agent session.