Frontier Red Team: Unreleased Claude Chained Flaws Into Working Exploits
Anthropic's Frontier Red Team says an unreleased research model built end-to-end exploit chains across five platforms. They include a zero-click remote-root Linux chain (CVSS 9.8), a one-click Chrome escape patched in Chrome 156, a Safari same-origin break, a locked Pixel 9 bypass over USB and a Windows QUIC crash.
All flaws were reported to vendors before patches shipped. The paper's argument: frontier models can now prove novel vulnerabilities with working exploits and little human help.
Anthropic Expands AI for Science With Free Credits
Anthropic is offering free Claude credits to 20 research foundations and institutions as part of its expanded AI for Science program, announced Oct. 28.
Gist: Claude Code Agent Summaries Eat 15-20% of Tokens
A GitHub gist reports that Claude Code's agent summary burns 15-20% of a user's tokens just to write short status lines. One reported case put it at 16.4% of weekly spend.
It's one user's measurement, not an Anthropic finding. If your weekly limit vanishes fast, it's worth a look. Latest release is still 2.1.311 (Oct. 26).
Research: Built-In Browsers in AI Coding Tools Can Be Broken
Forever Security's EMBRACE research reports flaws affecting the built-in browsers in Cursor, Copilot, Codex, VS Code and Claude Desktop. Details are thin in the summary, so check vendor advisories before drawing conclusions.
Status Page: All Systems Operational
The status page shows all systems operational. The most recent incidents were the Gmail connector (Oct. 27 23:00 to Oct. 28 05:29 UTC) and the Google Calendar connector (Oct. 26 to Oct. 28 18:52 UTC). Both are resolved, and no new incident is listed after them.
SpaceX's “Customer B” Is Anthropic, Per Bloomberg
A SpaceX filing names a buyer only as “Customer B,” 34.5% of Q3 revenue and bigger than the US government. Bloomberg matched it to Anthropic's S-1, filed Oct. 26. Neither company confirmed it on the record.
The deal reportedly runs up to $120.5 billion through 2029 for Nvidia GPU capacity at SpaceX's Memphis data centers, about $3.1 billion a month at full run rate from December. After initial periods ending in January and March 2027, either side can leave with 90 days' notice.
Wired: Contractor's Claude Agent Ran on a DOD Network for Two Months
Wired reports an Accenture staffer connected a personal Claude Code account to a secure Department of Defense system. The agent ran for about two months before detection and appears to have sent system details, including login credentials, to a high-risk third-party site.
The system holds Controlled Unclassified Information. The DOD declined to comment. Wired doesn't say how it was caught or which service got the credentials.
Nscale Pushes $30B IPO With $44.6B Anthropic Backlog
Bloomberg reports Nscale is pressing ahead with a $30 billion IPO despite investor pushback. A $44.6 billion Anthropic backlog is at stake. Compute demand from Anthropic keeps showing up in other companies' filings.
Anthropic's Growth Is Now Someone Else's Concentration Risk
Two stories, one theme: Claude is everywhere, and the edges are fraying. SpaceX now books a third of its quarter from one AI lab on a deal either side can exit with 90 days' notice. Anthropic's compute bills are becoming other companies' revenue risk.
Meanwhile an unsanctioned personal agent on a Pentagon network and a red team that can chain exploits show the same gap. Capability is outrunning governance. Expect enterprise buyers to ask who approved the agent, not just what it can do.