OpenAI shipped a purpose-built hacking model that found 400+ kernel zero-days before launch, Microsoft's August Patch Tuesday revealed Lazarus Group deploying a new FudModule rootkit via a WinSock zero-day, and an unpatched GeoServer SQL injection is already under active exploitation. Here's what mattered this week.
- OpenAI launched GPT-5.6-Cyber — a purpose-built offensive security model that found 2 Chrome zero-days and 400+ kernel privilege-escalation bugs before its release, available only to vetted researchers through Daybreak Red
- OpenAI also paused Astra — its next unreleased model may have crossed the "Critical" autonomy threshold under OpenAI's own Preparedness Framework
- Microsoft August Patch Tuesday: 421 CVEs, 3 zero-days — the actively exploited one (CVE-2026-68820, WinSock) is confirmed Lazarus Group, deploying a new FudModule kernel rootkit
- GeoServer zero-day RCE under active exploitation — unauthenticated SQL injection, no patch available, exploitation began within hours of public disclosure
- CEVA Logistics breach hits 8 European warehouses — Valve/Steam hardware customers, Dutch retailer Bol, and others had names, addresses, and order data stolen
- DeadLock ransomware disables Windows Defender, event logs, and backups before encrypting — a deliberate defense-evasion-first approach
- Water utility attacks spread further — New Jersey and Alabama join the growing list of targeted US states
- 737 malicious VPN/proxy Chrome extensions found routing users through attacker-controlled infrastructure
OpenAI ships a hacking model — and pauses the next one for being too capable
Two OpenAI announcements this week, and they tell opposite stories.
On August 10, OpenAI launched GPT-5.6-Cyber, a purpose-built offensive security model trained specifically for zero-day discovery and exploit-chain development. The headline number: it completes 95% of advanced offensive security prompts where standard GPT-5.6 Sol completes just 1.5% — a refusal metric, not a raw capability score, but still significant. Before launch, the model was used internally to find real vulnerabilities: two previously unknown Chrome V8 bugs (now patched by Google as CVE-2026-15903, CVSS 8.8, chainable to escape the V8 heap sandbox), five vulnerabilities in a popular mobile OS, three critical database flaws, and over 400 kernel privilege-escalation vulnerabilities in a popular OS. Access is tightly gated through Daybreak Red, requiring identity verification, legal attestations, and approved use cases — you can't reach it through the standard API or ChatGPT.
On August 7 — three days earlier — OpenAI separately announced it had paused development of Astra, its next unreleased flagship model, after internal evaluations indicated it may have crossed the "Critical" threshold on cybersecurity under OpenAI's Preparedness Framework — meaning the model could autonomously identify and exploit zero-day vulnerabilities in hardened real-world systems without human direction. GPT-5.6-Cyber was evaluated at "High" — one level below Critical — before release. So OpenAI shipped the model that stayed under the line and slowed down the one that may have crossed it.
Why both matter together: the same week OpenAI hands vetted defenders a model that can find hundreds of real zero-days, it's also quietly signaling that the next generation is already too dangerous to ship without more work. That's the tension the industry is navigating right now — and it's playing out in real time.
August Patch Tuesday: Lazarus Group, a wormable DNS bug, and 421 CVEs
Microsoft's August Patch Tuesday landed on August 11 with 421 CVEs — down from July's record 570, but still a heavy release. Three zero-days, one actively exploited:
CVE-2026-68820 (WinSock use-after-free, CVSS 7.0) — the must-patch item this month. Check Point confirmed that North Korean Lazarus Group exploited this flaw to deploy a new version of FudModule, its kernel-mode rootkit. A locally authenticated low-privileged user can race the driver to gain SYSTEM privileges — the kind of local privilege escalation that's a critical second-stage component in larger intrusions. This is the fourth afd.sys zero-day exploited in the wild since 2022, all linked to Lazarus activity.
CVE-2026-62832 (Windows User Profile Service, publicly disclosed) — an authenticated attacker with credentials for any local account can load another user's registry hive, gaining admin privileges. Matches the "LegacyHive" technique disclosed last month; Microsoft assesses exploitation as "More Likely."
CVE-2026-62878 (Windows DNS Server, CVSS 9.8) — a stack-based buffer overflow described by Zero Day Initiative as technically wormable. No authentication or user interaction required. Not yet exploited, but CVSS 9.8 and wormable in DNS infrastructure is a top-of-queue item regardless.
Beyond the zero-days: 62 Critical vulnerabilities total, 42% of the release is elevation-of-privilege flaws, and the release also covers Azure, Exchange, SharePoint, and GitHub Copilot. Adobe patched critical flaws in ColdFusion, Commerce, and Campaign Classic the same day.
GeoServer zero-day: unpatched, unauthenticated, already under exploitation
An unpatched SQL injection vulnerability in GeoServer — the open-source web server widely used in government, defense, science, and engineering for managing geospatial data — was publicly disclosed on August 12 by researcher @q1uf3ng on X, and exploitation attempts began within hours of public disclosure, with hundreds of attempts traced to a small pool of IP addresses. No CVE has been assigned yet and no patch is available.
The flaw sits in GeoServer's jsonArrayContains filter function — user-supplied arguments aren't properly sanitized before encoding into database queries, and under certain database configurations (particularly PostGIS and Oracle JDBC with sa database access), this leads straight to remote code execution. This one warrants immediate attention if you run GeoServer in internet-facing infrastructure, especially in government or research environments where it's most commonly deployed.
CEVA Logistics breach — Valve, Bol, and 8 European warehouses
CEVA Logistics confirmed to TechCrunch that a cyberattack lasting from July 29 to August 1 affected at least eight warehouses across Europe. CEVA is a wholly-owned subsidiary of French shipping giant CMA CGM ($18.3B revenue) and operates in 170+ countries. The blast radius is significant: affected clients include Valve/Steam, whose hardware customers in Europe had names, email and home addresses, phone numbers, and order details exposed, and Dutch online retailer Bol, whose operations at CEVA's Veerweg location faced extended restoration delays.
CEVA has not yet filed regulatory reports. The shutdown of physical warehouse operations (rather than just systems) points strongly toward ransomware or disruptive malware. It's worth noting CMA CGM was itself hit by ransomware in 2020 — this family of companies has now been in attackers' crosshairs for years.
DeadLock ransomware: defense-evasion first, encryption second
A newly tracked ransomware group has adopted a deliberate sequencing approach: before deploying encryption, DeadLock systematically disables Windows Defender, Volume Shadow Copies, backup services, and Windows Event Logs — removing the defender's ability to detect, recover, or investigate after the fact. It's a more methodical approach than the spray-and-encrypt operations of earlier ransomware generations, and signals increasing operational maturity even at the commodity end of the ransomware market.
CISA issued guidance this week urging organizations to patch exposed VPNs and segment networks in response to the growing Gunra ransomware campaign, separately.
Threat actor roundup
- Lazarus Group (North Korea) — actively exploiting CVE-2026-68820 with a new FudModule rootkit variant (see Patch Tuesday above)
- Head Mare — exploiting unpatched TrueConf server vulnerabilities to deliver trojanized installers inside trusted enterprise software channels, targeting Russian companies across instrumentation, energy, and transport sectors
- China-linked StormEncryptor ransomware — likely delivered via the N-central RMM flaw from two weeks ago, now being tracked as a distinct campaign
- Kimsuky (North Korea) — adopting offline AI tools to enhance attack operations, evading network-based detection by running models locally
Also worth a skim
- 737 malicious VPN/proxy Chrome extensions found routing traffic through attacker-controlled infrastructure, primarily targeting Russian-speaking users seeking access to blocked services
- Red Hat Kubernetes flaw allows privilege escalation to cluster-admin — patch promptly if you run OpenShift or RHEL-based Kubernetes
- VMware vCenter CVE-2026-59310 (CVSS 9.8, directory traversal → RCE) — already confirmed exploited in incident response cases, patched by Broadcom late July
- AI tool found 84 flaws in 5G network software, 23 of which remain unpatched
- DDoS attacks exceeding 1 Tbps surged fivefold in Q2 2026, per new research
- Levi Strauss is investigating unauthorized endpoint access and potential corporate data exfiltration
- EU data protection watchdog raised red flags this week over proposals to expand Europol's surveillance and data-sharing powers
- Abyssos RAT — a new remote access trojan hijacking live browser sessions, stealing credentials, and granting attackers VNC access, now circulating in the wild
Closing thought
This week had a theme hiding underneath the headlines: the AI offensive capability gap is now out in the open, and official. OpenAI's GPT-5.6-Cyber found over 400 real kernel bugs before it launched, and OpenAI itself pumped the brakes on the next model for being too dangerous to ship. Meanwhile Lazarus Group is exploiting a patched-today Windows zero-day with a kernel rootkit, GeoServer is under active attack with no fix available, and CEVA's breach shows again that your third-party logistics vendor is as much a part of your attack surface as your own perimeter. The scale and simultaneity of this week's stories is a useful reminder that threat actors aren't waiting for the industry to catch up.
Sources: The Hacker News, BleepingComputer, SecurityWeek, TechRepublic, Infosecurity Magazine, TechRadar, Qualys, CrowdStrike, eesel AI, TechTimes, CSO Online, Techmaniacs, Rankiteo, IT Security News.
Top comments (0)