U51

Episodes

Episode 22

All episodes
Episode 2211:56

Episode 22: OpenAI GPT-6 Cyber finds nearly a thousand Linux root paths, Metabase under active attack and ASD open-weight AI guidance, 6 November 2026

SpotifyApple PodcastsPocket Casts

OpenAI named GPT-6 Cyber as the model behind a defensive autonomous pipeline that reproduced 1,724 Linux kernel vulnerabilities and produced 948 working local privilege escalation exploits plus candidate patches in a mid-October run, at about US$1.2 million in API costs, with humans reviewing before maintainer reports. Singapore's CSA warns Metabase CVE-2026-72898 is under active exploitation, VulnCheck published a ten-flaw Control Web Panel chain to unauthenticated root, and ASD's ACSC released open-weight AI model guidance.

Australian focus

  • ASD's ACSC: new guidance on open-weight AI models (5 November 2026). Two publications: opportunities and cyber security risks, and selecting, deploying and operating open-weight models securely (governance, provenance, supply chain, monitoring, incident response). cyber.gov.au
  • ASD's ACSC with CISA and NSA: refreshed post-quantum questions to ask vendors (5 November 2026). cyber.gov.au

Global threats

  • Metabase unauthenticated SQL injection actively exploited (CVE-2026-72898, CVSS 10.0). First disclosed in August 2026 after a cloud zero-day attack (Metabase patched cloud then). Singapore CSA alert 4 November flags active exploitation now, especially unpatched self-hosted instances; PoC public; abuse of POST /api/session/reset_password can yield administrator access. Workaround: block that endpoint if upgrade is delayed. Post-upgrade: revoke sessions, review admins and API keys, rotate connected-database credentials, check logs. Minimum fixed releases include Open Source 0.58.24, 0.59.21, 0.60.17, 0.61.11, 0.62.9, 0.63.5 and matching Enterprise 1.x point releases (see Metabase advisory for the full matrix). Singapore CSA; Metabase security update; GitHub advisory GHSA-vwf4-m7j8-wcjf
  • Control Web Panel: ten flaws chain to unauthenticated root (VulnCheck, 3 November 2026). Entry CVE-2026-73695 (and related IDs including CVE-2026-73696, CVE-2026-73697, CVE-2026-77876 through CVE-2026-77881, CVE-2026-104080). VulnCheck Initial Access Intelligence customer/community exploit module plus detailed public write-up (not a free public PoC dump). Fixed in CWP 1.29; 1.30 current at disclosure. Prefer keeping ports 2083 / admin panels off the public internet. VulnCheck
  • CISA KEV: Apple CVE-2024-27834 (added 4 November 2026, federal due 18 November). Improper authentication / Pointer Authentication bypass; fixed May 2024 in Safari 17.5, iOS/iPadOS 17.5 and 16.7.8, macOS Sonoma 14.5, tvOS 17.5, watchOS 10.5. Press links the listing to the APT29-linked Poisoned Source campaign covered in Episode 21. CISA; The Hacker News
  • Apache OFBiz security release. All versions prior to 24.09.08 affected; includes pre-authentication RCE CVE-2026-68524 and many other CVEs. Upgrade to 24.09.08. oss-security; OFBiz security
  • DreamHost File Manager incident. Unauthorized access discovered 23 October 2026; FTP/SFTP credentials of File Manager users (about August to 23 October) and a limited set of webmail credentials exposed; passwords reset; no evidence of control-panel or payment-card access per DreamHost. DreamHost

AI security

  • OpenAI GPT-6 Cyber / A thousand fewer paths to root. Pipeline note 4 November 2026; GPT-6 Cyber announcement ~early morning AEDT 6 November (late item after 5:45am draft). Mid-October autonomous pipeline powered by GPT-6 Cyber: 2,308 candidate findings, 1,724 reproduced, 948 distinct Linux kernel local privilege escalation exploits with candidate patches; about US$1,213,430 in API credits (about US$1,279 per LPE). Humans review before maintainer reports. 304 critical findings disclosed with candidate patches so far (not all 948). Within 10 days, 61 LPEs already patched upstream by other contributors; 43 staging; about 299 discussed publicly. Ten new techniques in a sample of 166 exploits checked (details withheld). Access initially limited to selected approved Daybreak Red customers via the API and Codex (identity verification, SSO, MFA, managed devices, IP allowlisting); expanding to all Daybreak Red customers in coming weeks. US$1 billion Daybreak for Frontline Defenders grant programme for resource-constrained teams (government, essential services). OpenAI rates the model Critical for cyber under its Preparedness Framework. OpenAI also says the model found 81 new high-severity Firefox bugs (Mozilla's rating); Mozilla has begun shipping fixes. OpenAI thousand paths; OpenAI GPT-6 Cyber; The Hacker News; SecurityWeek
  • UK NCSC: Artificial intelligence and cyber intrusion assessment. Baseline judgment that AI-powered tools have almost certainly increased zero-day discovery; defensive AI can shorten zero-day to n-day windows if configured and matched by faster patch processes. NCSC PDF
  • Japan seeking earlier access to undisclosed vulnerability data from US AI firms (Nikkei via iTnews, 5 November 2026). Emergency cybersecurity package may go to ministers as early as 6 November. Single-source lead. iTnews

Outlook and patch priorities

  • Critical: Metabase self-hosted vulnerable builds (exploited); Control Web Panel 1.29 or later; Apache OFBiz 24.09.08.
  • High: Apple builds at or above the May 2024 fixed releases above; current Linux kernels and distributions; DreamHost File Manager users rotating FTP/SFTP and reused passwords.
  • KEV in window: CVE-2024-27834 (Apple), added 4 November 2026, due 18 November 2026.