Saturday, October 24, 2026

TL;DR

Patch self-managed GitLab now - a CVSS 10.0 flaw and a Railway secrets breach landed Oct 21, feeding what Zvi Mowshowitz calls AI-accelerated attacks outpacing patches. METR found real monitoring gaps at both OpenAI and Anthropic this week. OpenAI cut its funding-round valuation target to $1.25 trillion from $1.4 trillion on weak investor demand.

Act on this

  • Patch self-managed GitLab now: CVE-2026-110264, CVSS 10.0, unauthenticated code execution since 2014. Railway found attacker access to customer env vars - rotate secrets. CSA Singapore advisory #

Signals

AI is compressing the time from a flaw being found to being exploited at scale faster than defenders can patch. #

Zvi Mowshowitz's Oct 23 'The Hackening' ties together one week: a new unauthenticated GitLab flaw (CVE-2026-110264, CVSS 10.0) fixed Oct 21 while Railway reported attacker access to customers' encrypted environment variables; a Chinese hacker breached 100+ companies, including Deere and SharkNinja, for about $50 a breach using AI agents, per Reuters; and major US banks began detecting unauthorized AI agents inside customer accounts, per Axios. Zvi Mowshowitz, The Hackening

Independent auditors are finding concrete monitoring failures inside frontier labs' own production systems, not just publishing incident postmortems after the fact.  #

METR's Oct 20 review of Anthropic's August Risk Report found the offline monitor that flagged just 1% of a scheming Mythos 5's actions jumped to roughly 50% once its chain-of-thought was hidden - meaning production monitoring may already miss reasoning-based rationalization. Two days later, a joint METR-Redwood audit inside OpenAI found working bypasses of its Codex sandbox and monitor; OpenAI says the production-affecting ones are now patched. METR, review of Anthropic's Risk Report · METR, red-teaming OpenAI's agent monitoring

America's open-weight models are closing the gap with China's, not just continuing to trail.  #

Nathan Lambert's Oct 17 Interconnects review says Meta's 816B Muse Spark Open and Reflection's Beam have roughly halved the capability gap to the best Chinese open models, GLM 5.3 and Kimi K3 - without them, Lambert counts 20+ Chinese open-weight models beating the best Western ones. He still flags Meta's custom license, which requires attribution and a paid tier above $20 million in monthly revenue, as a strategic risk. Nathan Lambert, Meta shakes up the frontier's open

News

WATCHOpenAI cuts its funding-round target to $1.25 trillion, down from $1.4 trillion #

The Financial Times reports OpenAI is now seeking $30 billion near a $1.25 trillion valuation, after investors called demand a 'thud' against past rounds. UAE investors and Coatue are said to lead; terms aren't final, with an October 30 commitment deadline. CNBC, OpenAI funding round

ACTMicrosoft traces a China-linked attack using an AI pentesting tool to hundreds of breached organizations  #

Microsoft says Storm-0940 used the open-source ARTEX agentic pentesting framework, letting operators plug in any AI model to automate reconnaissance and exfiltration, reaching privileged access in under three hours at some targets via compromised ConnectWise Cloud Backup accounts. Reuters, via KFGO

WATCHMicron's Taichung union in Taiwan settles; the Taoyuan strike's outcome remains unclear  #

Micron Memory Taiwan and its Taichung union signed a labor agreement Oct 22 and withdrew from mediation, terms undisclosed. The separate Taoyuan union's one-day warning strike, backed by a 99% authorization vote, was due the next day; whether it went ahead is unconfirmed. Taipei Times, Micron union Taiwan

WATCHNvidia strikes deals with six suppliers to expand AI chip packaging capacity  #

Nvidia announced deals at GTC Berlin with Amkor, Samsung Electro-Mechanics, SPIL and others to roughly triple interposer wafer and substrate capacity by 2028, plus a $1.1 billion advance to KYEC for US chip-test capacity - confirming packaging, not chip design, as the binding constraint. Nvidia, supply chain capacity expansion

WATCHGAO finds the Pentagon's AI oversight office still has just 3 of 10 positions filled #

A GAO report found the Defense Department's Responsible AI Division had only 3 of 10 authorized positions filled as of February, with 29 of 64 strategy items incomplete; its oversight council reportedly didn't meet from late 2024 until August 2025. Military Times, Pentagon AI oversight

The long view

A year ago, OpenAI's valuation hit $500 billion through an employee share sale alone, making it the world's most valuable private company, with investor demand strong enough that no persuasion was needed. On October 22, the Financial Times reported OpenAI cutting its target for a fresh $30 billion round to about $1.25 trillion, down from $1.4 trillion, after investors called demand a 'thud' versus past raises - the round explicitly timed to close before Anthropic's IPO. If this holds, OpenAI's fundraising now depends on beating a rival to market, not investors competing to get in.

Also noted

  • Claude Code v2.1.310 extended its Monitor tool to Bedrock, Vertex and Foundry and fixed a long-standing compaction bug. Claude Code releases  #
  • Zvi Mowshowitz reports Indian certificate authority eMudhra had 8 private keys compromised; one was guessable as just the digit 4. Zvi Mowshowitz, The Hackening #