IT News Roundup: SonicWall Exploited in the Wild, Mistral Large 4, GhostAction Supply-Chain Attack — Week of Oct 5 – Oct 11, 2026

The week belonged to edge appliances under fire and to the open-weight frontier moving fast. A maximum-severity SSRF flaw in SonicWall SMA1000 gateways, patched only three days before this roundup, is already being exploited in the wild, and Citrix followed with an urgent patch warning for a critical NetScaler RCE. The GhostAction supply-chain campaign returned to GitHub, planting credential-stealing workflows in tens of thousands of repositories, while a working root exploit for an unassigned AnyDesk Linux flaw went public. On the model side, Mistral previewed its 1-trillion-parameter open-weight Large 4, Reflection announced the 501B Beam, and llama.cpp shipped first-class support for decision models. OpenAI meanwhile confirmed it is raising fresh capital at a valuation near $1.4 trillion.

Deep Dive: SonicWall SMA1000 Flaw Patched Tuesday, Exploited by Friday

Attackers are actively exploiting a maximum-severity vulnerability in SonicWall SMA1000 appliances, tracked as CVE-2026-102255, that SonicWall patched on Tuesday, October 6. The flaw affects the Appliance WorkPlace interface on SMA1000 6210, 7210, and 8200v models; the SMA 100 Series and SSL-VPN running on SonicWall firewalls are not affected. According to SonicWall's advisory, a remote unauthenticated attacker can abuse the interface to direct the appliance to issue requests on the attacker's behalf, reach internal functionality, and perform unauthorized operations. Source: BleepingComputer.

The exploitation evidence comes from the Previdian honeypot network, which detected requests consistent with the flaw on Friday, October 9. The crafted requests target the WorkPlace Extraweb interface with an OPTIONS request that reaches the appliance's internal CouchDB service at 127.0.0.1:5984, then attempts to traverse into a CouchDB design document and invoke its _rewrite function while supplying an HTTP Basic Authorization header carrying the credentials admin:admin. The technique differs from the SSRF flaws disclosed against the same WorkPlace interface in July and September 2026, but the target surface is the same. Source: BleepingComputer.

Context matters: the SMA1000 line is a perennial target because managed service providers, large enterprises, and government agencies use these secure remote access gateways for VPN access to internal applications. In July, threat actors abused two SMA1000 zero-days (CVE-2026-15409 and CVE-2026-15410) for weeks to install custom Sou5, OrangeTail, and RootRun malware, and CISA later linked some of those attacks to ransomware gangs. Last month, attackers chained two further zero-days (CVE-2026-83548 and CVE-2026-83549) in the wild to execute remote code on vulnerable gateways. Over the past four years, CISA has added 19 SonicWall vulnerabilities to its Known Exploited Vulnerabilities catalog, flagging 13 of them as used by ransomware operations.

Shadowserver currently tracks more than 400 SMA1000 appliances exposed on the internet, though there is no information on how many are honeypots or already patched. SonicWall had not yet flagged the vulnerability as actively exploited in its Tuesday advisory; the in-the-wild signal arrives from third-party monitoring. That gap between patch Tuesday and confirmed exploitation is the window in which exposed appliances are most at risk.

For IT professionals and homelabbers: update SMA1000 6210, 7210, and 8200v devices to the patched release immediately, verify whether the WorkPlace interface is reachable from the internet at all, and review logs for anomalous OPTIONS requests or unexpected internal traffic toward CouchDB endpoints. If an appliance cannot be patched right away, restricting access to the WorkPlace interface and monitoring for the admin:admin Basic auth pattern are reasonable interim controls. Given the appliance's role in MSSP and government VPN stacks, treat any unpatched internet-facing SMA1000 as presumptively exposed.

AI Models & Weights

Mistral Large 4: A 1-Trillion-Parameter Open-Weight Frontier, Weights Due by Month's End

Mistral launched a public preview of Mistral Large 4 on October 6, unofficially ML4, officially le Chonk. It is a 1-trillion-parameter natively multimodal mixture-of-experts model with 52 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. The preview is served on that same infrastructure via the Mistral Studio API, and the weights are promised by the end of October. Source: Mistral.

Mistral claims the model is competitive with the strongest open-source models globally and outperforms any open-weight model developed in the US or Europe, with state-of-the-art results among open models on enterprise workloads in cybersecurity, finance, and law. On the Artificial Analysis Cyber Index it ranks among the top five models globally and leads open-weight models developed outside China by a wide margin, scoring 82% on a test that asks a model to reproduce a real open-source vulnerability and then patch it, the highest of any model, and solving 93% of Cybench challenges. Notably, several leading closed models score near zero on that same test because they refuse to perform it. A significant share of training data was multilingual, spanning more than 160 languages including every official EU language.

For local use, the practical note is timing: no weights, architecture details, or license are public yet, and the model is reachable only through a guardrailed Studio preview. At 1T total parameters with 52B active, expect a multi-GPU serving footprint once the weights land, and watch for GGUF builds and the stated license. For security teams, the headline is that a self-deployable open model now claims frontier-class offensive-defensive cyber capability, which changes the calculus for running vulnerability research and incident response under your own policies.

Reflection Beam: 501B Open-Weight MoE, 1M Context, Apache 2.0 Weights Due This Month

Reflection, the Nvidia-backed startup founded by two former Google DeepMind researchers, introduced Beam on October 5, its first open-weight model. It is a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active, a 1-million-token context, and an Apache 2.0 license promised for the weights, technical report, and model card later this month. Source: Reflection.

On Reflection's own benchmark table, Beam sits level with GLM 5.2 and behind GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash on most coding and agentic tests: 44.4 on DeepSWE v1.1, 80.1 on Terminal Bench v2.1, and 80.9 on SWE-Bench Verified. The company's argument is efficiency per token rather than leaderboard position, estimating 3 to 4 times less inference compute than GLM 5.2 at comparable scores. The training run was substantial: 23.8 trillion tokens across 6,144 GB300 GPUs in under four weeks, followed by more than 100 million reinforcement-learning rollouts on 10.5K GB300 GPUs across roughly one million curated environments.

All figures are vendor-reported; no independent reruns exist until the weights drop. The architecture uses interleaved local and global attention with fine-grained routed experts across 52 layers, and Reflection reports near-uniform expert utilization, with the busiest expert at 1.04 times average load. For homelabbers, the 1M context and Apache 2.0 license are the hooks; the 23B active parameter count suggests a serving profile far friendlier than the 501B headline implies, but wait for the actual checkpoint and community quantizations before planning hardware.

llama.cpp Ships First-Class Decision Models

The ggml-org team added native decision-model support to llama.cpp, a new /v1/systemone endpoint that answers structured questions in a single forward pass instead of generating free-form text. Three question types are supported: choice (returns the top option plus a probability per option), score (an expected value on a 2 to 10 level scale), and noul (a yes/no probability). A request carries a state plus one or more questions, and the server answers them independently in one pass. Source: Hugging Face.

The launch set includes models from 0.1B to 27B parameters in GGUF, such as Kev-4B, lev, OpenJev, and the 27B Clef, with the larger models answering a single question in 36 to 43 milliseconds on an NVIDIA RTX PRO 6000. OpenJev is multimodal and can read documents or screenshots, with the vision projector downloading automatically. Router mode lets one server host several models and load them on demand, picking a model per request.

For local inference setups, decision models are a concrete way to offload high-volume structured tasks, ticket routing, intent classification, triage, onto cheap local hardware without burning frontier API tokens. The practical pattern is to act on confident answers and route the rest to a human, choosing the confidence cutoff per model after testing on your own examples.

The Rest of the Week

Citrix: Patch a Critical NetScaler RCE Immediately

Citrix warned administrators on October 9 to patch systems immediately against CVE-2026-107406, a memory-overflow vulnerability in NetScaler ADC and NetScaler Gateway that can lead to remote code execution or a denial-of-service crash. To be vulnerable, the appliance must be configured as a SAML Identity Provider or Service Provider. Citrix is not aware of any unmitigated exploitation at the time of the bulletin, but the company urged upgrades to NetScaler ADC and Gateway 14.1-73.46 and later, or 13.1-64.29 and later, with separate FIPS and NDcPP targets. Source: BleepingComputer.

The exposure is large: Shadowserver tracks over 21,000 IP addresses with NetScaler fingerprints exposed on the internet, including just over 1,500 Gateway instances and nearly 20,000 ADC appliances. Citrix has had a string of exploited NetScaler zero-days this year, including CVE-2026-3055 and CVE-2026-4368 in March and the actively exploited RCE zero-days CVE-2026-88771 and CVE-2026-88772 in September, and CISA has flagged 27 actively exploited Citrix vulnerabilities since November 2021, seven of them used in ransomware attacks. The pattern of exploitation following disclosure within days makes prompt patching the only reliable control.

Teams should inventory which appliances run SAML IdP or SP configurations, verify firmware levels, and patch on the expedited track. If exposure is internet-facing, treat the window between this bulletin and patching as high risk.

GhostAction Returns: Credential-Stealing Workflows in Tens of Thousands of GitHub Repositories

Security researchers disclosed an ongoing credential-theft campaign that compromised two high-profile open-source maintainer accounts to push a malicious GitHub Actions workflow into over 340 repositories. Using the account of Takashi Kitao, author of the 18,400-star pyxel game engine, the attacker pushed the workflow to 27 repositories; eight hours later, the account of Henry Wu, the original author of Uber's athenadriver, was used to push the same workflow to 318 repositories in a 16-minute window. As of October 9, Socket had identified more than 500 GitHub accounts that committed the malicious workflow to tens of thousands of repositories since October 7. Source: The Hacker News.

The activity is attributed to GhostAction, a supply-chain campaign first documented in September 2025 that stole 3,325 secrets across 817 repositories. The injected workflow, named security-audit.yml or github_actions_security.yml, triggers on workflow_dispatch and any push, checks out with fetch-depth: 0, and exfiltrates the repository's named secrets plus 13 credential patterns, AWS keys, Anthropic, OpenAI, and OpenRouter API keys, and GitHub and GitLab tokens, to a hard-coded attacker IP over plain HTTP. It scans the working tree and the entire git history, so previously committed and deleted credentials are harvested too.

The most exposed targets are private forks and downstream mirrors, because private repositories are where committed credentials actually live; every run also returns a repository identifier, giving the operator a map of reachable execution contexts. Developers are advised to check repositories for either workflow file since August 31, assume compromise if present, revoke the compromised credential, rotate all affected secrets, delete the workflow from all branches, and audit forks. Source: The Hacker News.

Anthropic Launches a Free AI Vulnerability Scanner for Open-Source Projects

Anthropic unveiled OSS Scanner on Thursday as part of a broader Cyber Mission program to protect critical infrastructure and open-source software. The opt-in service runs thorough, periodic security scans of participating projects using the company's strongest models, automatically flagging and explaining vulnerabilities and suggesting patches. Outputs are fully model-generated and ship without human review or triage, which Anthropic says enables faster and more frequent scanning. Source: The Decoder.

Anthropic expects accuracy above 90 percent, but the company is explicit that reports may contain errors. The scanner is informed by the company's experience using Claude to find vulnerabilities during Project Glasswing, and the argument is that attackers already have powerful models hunting for bugs while defenders lack comparable tooling. Maintainers of projects critical to infrastructure or user safety can opt in via GitHub. Source: Anthropic.

The companion Critical Infrastructure Defense Program gives operators of power grids, water systems, and transportation networks access to Claude models, engineers, and threat analysis, with CrowdStrike, Palo Alto Networks, Deloitte, and Rockwell Automation as founding partners. For open-source maintainers, the trade-off is clear: free, frequent, AI-driven triage in exchange for accepting unreviewed model output into the project's security workflow.

OpenAI Seeks $30 Billion at a Valuation Near $1.4 Trillion

OpenAI is negotiating at least $30 billion in new capital at a target pre-money valuation of $1.4 trillion, according to Bloomberg reporting on October 9. The company's annualized revenue rate sits at about $50 billion, a correction from an earlier reported $70 billion figure that used a different accounting method, and that correction sent chip stocks sliding. OpenAI expects to actually reach at least $70 billion in annualized revenue by the end of 2026, driven by its expanding enterprise business, with overall annualized revenue up 77 percent in Q3 and enterprise revenue up 107 percent. Source: The Decoder.

The company has pushed its IPO to 2027, and the round lands amid the broader AI-bubble debate over whether revenue growth can keep pace with massive compute spending. At DevDay, OpenAI reported more than 1.2 billion weekly ChatGPT users, over 35 million weekly ChatGPT Work and Codex users, and 2.5 million businesses on its products. The funding round follows a March raise of up to $122 billion at a post-money valuation of $852 billion.

For IT professionals, the through-line is enterprise adoption: the revenue growth is being driven by business customers rather than consumer subscriptions, which is where the platform commitments, data residency questions, and integration work actually land.

Microsoft: Outdated Windows Devices Lose Security Updates in 2027

Microsoft warned that outdated versions of Windows will lose access to Windows Update services and stop receiving security updates entirely once it rotates its Windows Update signing certificates in May and June 2027. The company recommends upgrading to a supported version and published version-specific guidance: Windows 11 25H2 and later require no action; Windows 11 24H2 and Windows Server 2025 must install the September 2025 security update or later before June 19, 2027; other supported versions of Windows 11, Windows Server 2022, and Windows 10 must install the July 2026 update or later before June 19, 2027; and Windows 10 Enterprise 2019 LTSC, Windows Server 2019, and Windows Server 2016 must install the July 2026 update or later before May 17, 2027. Source: BleepingComputer.

Replacement certificates have already been delivered to systems running a supported Windows version, so up-to-date machines need no action. Devices receiving updates through WSUS are handled separately, since WSUS downloads and redistributes updates across the enterprise network. The practical step is to inventory every device running an older or unsupported Windows version now, deploy monthly updates on supported devices, and build an upgrade plan for the rest before the May and June 2027 certificate expiration dates. Microsoft noted that Windows 11 26H2, the 2026 Update, is rolling out as a small enablement package to eligible 24H2 and 25H2 systems, which can simplify part of the upgrade path.

For homelab and small-office environments, this is a good prompt to find the forgotten Windows 10 and Server 2016 machines that still hold production data. Once the certificates rotate, those devices will silently stop receiving patches, which is exactly the kind of gap that gets exploited later.

TP-Link Sued by Four More U.S. States

Four more U.S. states sued router maker TP-Link Systems on October 6, bringing the total to five after Texas filed in February. Florida, Iowa, Montana, and Nebraska allege the California-based company misled buyers about how secure its routers are and how separate it is from China. TP-Link denies the claims and says it will fight them in court. Source: The Hacker News.

Until a 2024 restructuring, TP-Link Systems was affiliated with TP-Link Technologies, a Chinese company the suits do not name as a defendant. The complaints from Florida, Montana, and Nebraska do not allege that the Chinese government has obtained customer data through TP-Link; they describe that as a risk under Chinese law and note that state-backed hackers have exploited flaws in TP-Link routers. Iowa's filing is worded more strongly, with Attorney General Brenna Bird's office claiming TP-Link firmware gives the Chinese government access. The suits add legal pressure to an ongoing debate over whether consumer and SMB routers from Chinese-owned manufacturers belong in security-sensitive networks.

For IT professionals, the litigation does not change the immediate technical guidance, which is to keep firmware current, disable remote management, and treat consumer-grade equipment accordingly in any environment handling sensitive data. The case is worth watching for how the courts treat the relationship between the California entity and its Chinese parent.

Week in Numbers

• 400+ SonicWall SMA1000 appliances tracked as exposed on the internet by Shadowserver, three days after CVE-2026-102255 was patched and before confirmed in-the-wild exploitation.
• 19 SonicWall vulnerabilities sit in CISA's Known Exploited Vulnerabilities catalog over the past four years, 13 of them flagged as used by ransomware gangs.
• 500+ GitHub accounts identified by Socket as pushing the malicious GhostAction workflow to tens of thousands of repositories since October 7; the 2025 campaign stole 3,325 secrets.
• Mistral Large 4 is a 1-trillion-parameter model with 52 billion active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs; Reflection Beam is 501B total / 23B active with a 1M-token context.
• OpenAI is raising at least $30 billion at a target pre-money valuation of $1.4 trillion, with annualized revenue of about $50 billion and 1.2 billion weekly ChatGPT users.