IT News Roundup: WordPress RCE Exploited in Hours, GPT-6 Sol and Opus 5.5 Land — Week of Sept 21 – Sept 27, 2026

The week of September 21–27 was defined by speed: on the security side, a critical WordPress flaw moved from disclosure to active exploitation within hours, while CISA packed its Known Exploited Vulnerabilities catalog with enterprise-grade targets and set patch deadlines that land this weekend. On the AI side, OpenAI and Anthropic both shipped new models on the same day — GPT-6 Sol/Luna and Claude Opus 5.5 — Xiaomi released an open-weight omnimodal flagship under a permissive license, and Ollama made MLX the default runtime path on Apple Silicon. Rounding out the week: a $351.6 million cryptocurrency exchange heist attributed to North Korean actors, a precautionary six-hour shutdown advisory from Kiteworks, and an AI-agent-powered botnet hunting exposed Docker daemons.

Deep Dive: WordPress CVE-2026-87902 — Critical Path Traversal Exploited Within Hours

Threat actors began actively exploiting a critical security flaw in WordPress core within hours of its public disclosure on September 23. The vulnerability, tracked as CVE-2026-87902 with a CVSS score of 9.2, allows an unauthenticated attacker to achieve remote code execution (RCE) against affected installations — no login, no existing account, and no user interaction required.

The flaw lives in WordPress's page-template resolution logic. An unauthenticated attacker can manipulate the get_page_template() function so that it includes a chosen readable local .php file from outside the active theme directory, effectively turning path traversal into arbitrary code execution once the request reaches server-side template processing. Exploitation requires specific preconditions — a particular server environment and theme configuration must be in place — which is why not every installation was compromised, but security firms recorded numerous successful attacks shortly after disclosure.

In observed attacks, threat actors used files such as pearcmd.php to write malicious scripts to disk. PearCMD is a legitimate PHP command-line tool for managing PEAR packages, so its presence in a web root — especially one that was never intentionally deployed there — is a strong indicator of compromise. Researchers published a set of attacker IP addresses associated with the exploitation wave, which defenders can use to search access logs.

The affected range spans WordPress 4.7 through 7.1.1. Patches shipped as version 7.1.2, with backports for older supported branches: 7.0.6, 6.9.9, and 6.8.10. Default automatic security updates likely limited the number of successful compromises on stock installations, but any site that disables auto-updates — a common choice in managed environments to avoid unexpected changes — was exposed during the window between disclosure and patching.

For IT professionals and homelab operators, the practical steps are straightforward: update every WordPress installation immediately (including staging boxes and forgotten NAS-installed instances), audit web roots for unexpected .php files such as pearcmd.php, review access logs from September 23–24 against the published attacker IP list, and enable automatic security updates where change-management policy allows. The exploitation window measured in hours rather than days — a pattern that has become standard for critical WordPress flaws this year.

Sources: The Hacker News and CyberNews Weekly, Week 39 review.

AI Models & Weights

OpenAI Ships GPT-6 Sol and GPT-6 Luna at Half the Previous Generation's Price

On September 22, OpenAI released GPT-6 Sol and GPT-6 Luna, the mid-tier models of its GPT-6 generation, following the flagship GPT-6 Astra. OpenAI positions Sol as "built to power complex coding and agentic workflows." The model carries a 1,050,000-token context window, 128,000 maximum output tokens, and a knowledge cutoff of April 20, 2026.

The headline number is price: Sol is listed at roughly half the GPT-5.6 Sol rate, which OpenAI attributes to improved caching and inference — including higher default prompt-cache hit rates where changing reasoning effort or tools no longer invalidates cached context, plus a 90% discount on cached input reads. Function calling is supported on both the Responses and Chat Completions endpoints (the latter with an explicit flag). Rollout began September 22 across the API, Codex, ChatGPT Work for paid tiers, and the desktop app.

Both models are closed-weight, so there is no local-inference angle; the relevance for self-hosters is competitive pressure on API pricing as agentic workloads become a primary cost center. Teams migrating from GPT-5.6 Sol should note that "Sol" now spans two generations with different knowledge cutoffs — pin exact model identifiers in production configurations.

Sources: OpenAI release notes (via Releasebot) and EvoLink GPT-6 Sol release analysis.

Anthropic's Claude Opus 5.5 Matches the Flagship at a 40% Lower Running Cost

Also on September 22, Anthropic released Claude Opus 5.5, the first model in its new Claude 5.5 family. Anthropic states it performs at the level of its Fable 5.1 flagship on most work while costing about 40% less to run than Opus 5 on typical workloads at default settings.

Benchmarks back the claim: at medium effort, Opus 5.5 scores 54.6% on FrontierCode — ahead of GPT-6 Astra's top score of 53.3% at roughly a fifth of the cost per task — and leads Terminal-Bench 4.0 at extra-high effort. API pricing drops to $4/$20 per million tokens, with cache reads falling 60% to $0.20; output generation runs more than 30% faster than Opus 5. Cybersecurity and biology requests hit Fable 5.1-class safeguards, with most offensive cyber tasks re-routed to a lower-capability model.

Like the rest of Claude's lineup, Opus 5.5 is closed-weight and available via the Claude Platform, AWS, Google Cloud, and Azure — no self-hosting option. For API users, the cost-adjusted results are the story: agentic coding sessions that were priced around flagship rates now land at mid-tier economics.

Sources: Anthropic announcement and MarkTechPost benchmark breakdown.

Xiaomi Releases MiMo-V2.6-Pro-RL: Open-Weight, Omnimodal, 1M-Token Context

Xiaomi published MiMo-V2.6-Pro-RL on September 21 under the MIT license — one of the most permissive terms in the open-weight space. The flagship checkpoint of the MiMo-V2.6 series is natively omnimodal: text, image, video, and audio in a single model, with a 1M-token context window aimed at long repositories, tool traces, and multi-session agent runs.

The technical report emphasizes scaling reinforcement learning toward self-improvement — fully asynchronous GRPO on very large batches (1,568 prompts × 16 rollouts per step) combined with an agentic grader that compares rollouts within each group rather than relying on binary pass/fail rewards. The model trained in a single mixed RL run across coding, general agents, visual tasks, and cybersecurity.

The release quickly climbed Hugging Face's trending list. For local use, the full-precision flagship checkpoint is large; watch for community GGUF quantizations before committing VRAM budgets. The MIT license makes commercial self-hosting straightforward — no attribution or revenue-sharing strings attached.

Sources: Hugging Face model card and Xiaomi MiMo blog.

Ollama v0.34.4 Makes MLX the Default on Apple Silicon, Fixes Speculative-Decoding Memory Growth

Ollama shipped v0.34.4 on September 23 (with v0.40.0-rc0 following on September 25), and the release is significant for anyone running local models on Macs: model architectures supported by the MLX runtime now run on MLX by default on Apple Silicon, with more architectures being enabled during pre-release testing.

The update also exposes each model's thinking controls and defaults through /api/show, adds Nemotron H vision support via MLX, applies structured outputs in a single pass for thinking models (faster and more reliable), and speeds up Qwen 3.8 prompt processing with gated-delta and SwiGLU optimizations.

The most practically important fix targets memory: the MLX runner previously leaked freed KV buffers during speculative decoding, so long generations at large context could balloon past 90 GB on a 128 GB machine until the kernel panicked. The fix releases buffer pools every 256-token boundary — qwen3.8:27b-mlx at a 98k-token context now stays flat around 30 GB. Homelabbers running long-context local inference on Apple Silicon hardware should update.

Sources: Ollama GitHub releases and Releasebot Ollama notes.

The Rest of the Week

Kiteworks Urges Customers to Shut Down Servers for Six Hours Over Imminent-Attack Warning

Secure file-sharing vendor Kiteworks emailed customers on September 25 warning that it had received "credible threat intelligence from law enforcement indicating an attack on Kiteworks systems may be imminent this weekend." The company recommended a six-hour shutdown window on Saturday, September 26 — 4:00–10:00 a.m. in Central Europe, for example — and advised taking systems offline even if they are not directly reachable from the internet.

Kiteworks stressed that the advisory is precautionary rather than a response to a confirmed breach: no compromise of its systems was known at the time of the warning, and all currently identified vulnerabilities are addressed in release 9.5.1. However, Heise reported that Kiteworks customer support told the publication the shutdown recommendation exists specifically to protect against potential zero-day attacks — a distinction the official statement does not confirm.

The warning matters because Kiteworks' secure file-transfer and communications products are used by government organizations, financial institutions, and enterprises — platforms that store sensitive documents make them prime targets for data-theft extortion operations. The Clop gang has a long history of targeting enterprise sharing platforms in such attacks. Organizations running Kiteworks should confirm they applied the shutdown window, verify they are on 9.5.1, and monitor for follow-up advisories.

Source: BleepingComputer.

Suspected North Korean Hackers Steal $351.6 Million from Bitget

Cryptocurrency exchange Bitget disclosed on September 25 that suspected North Korean hackers stole $351.6 million from its hot and warm wallets — one of the largest crypto heists of the year. The breach was discovered Thursday evening after security systems flagged multiple unauthorized transfers originating from a limited number of wallets.

According to CEO Gracy Chen, the attackers compromised a critical backend system within Bitget's wallet infrastructure, used it to spoof transaction data, and triggered the exchange's authorization-signing process to move funds out. The theft spanned Ethereum, XRP Ledger, Arbitrum, Avalanche, Optimism, BSC, and Base chains, affecting ETH, XRP (the largest single-chain loss), BNB, AVAX, USDT, USDC, and other tokens. Bitget suspended all withdrawals while investigating with law enforcement, Mandiant, and SlowMist; some attacker wallet addresses have reportedly been frozen.

The exchange said its self-custodial Bitget Wallet was unaffected — it runs on independent infrastructure — and that the User Protection Fund, holding 5,500 BTC (roughly $464 million), will cover all losses. The attribution to North Korean groups follows IP behavior patterns and on-chain analysis consistent with known state-sponsored campaigns; Elliptic estimated in early 2025 that North Korean hackers have stolen over $6 billion in crypto since 2017.

Source: BleepingComputer.

CISA Adds WSO2, Adobe Commerce, SharePoint, and Mikrotik Flaws to KEV With Weekend Deadlines

CISA's September 25 update to the Known Exploited Vulnerabilities catalog added four actively exploited vulnerabilities. The two critical entries carry a federal patch deadline of Sunday, September 27: CVE-2026-5430, an authentication bypass in WSO2 products stemming from its JWT mechanism accepting tokens signed with an unsupported algorithm (affecting API Manager 4.1.0–4.6.0 and API Control Plane, Traffic Manager, and Universal Gateway 4.5.0/4.6.0), and CVE-2026-71362 (CVSS 9.1), an incorrect authorization flaw in Adobe Commerce/Magento that requires no existing account, administrator privileges, or user interaction.

The WSO2 entry is notable for its blast radius: the vendor's technology runs at nearly 1,000 customers across banking, government, telecommunications, and logistics. Security firm watchTowr said its honeypots captured exploitation attempts on September 13 using forged JWT tokens — though the attacker initially targeted the wrong product — and reproduced the attack successfully against the correct one, where a forged token could expose API endpoints and application credentials.

CISA also listed CVE-2026-65660, a high-severity code injection flaw in Microsoft SharePoint, and CVE-2026-67279, a medium-severity pre-authentication SSH state-machine bypass in Mikrotik RouterOS; federal agencies have until Monday, September 28 for those two. For any organization running WSO2 gateways or Magento storefronts — common in both enterprise and homelab edge stacks — the practical guidance is to patch before the weekend deadline rather than after confirmed exploitation.

Source: BleepingComputer.

Carbonato Botnet Uses an AI Agent to Hijack Exposed Docker Daemons

Researchers at ThreatDown documented a new worm-like botnet called Carbonato that targets hosts with unauthenticated Docker daemon APIs exposed on port 2375. The malware was discovered in an unauthenticated public registry containing nearly 60 repositories and 4.3 GB of image data, with operational evidence spanning October 2024 to August 2026.

The infection chain is efficient: Carbonato connects to the exposed Docker API, instructs the daemon to launch a privileged container (gaining host access), opens a reverse SSH tunnel, installs an SSH server carrying the operators' key, and establishes persistence through cron jobs, systemd timers, rc.local, and OpenRC hooks. The distinctive element is its control layer — the implant installs the Hermes Agent AI framework under an agent persona named "GH0ST," overwriting the default SOUL.md file. An operator drives it through Telegram: the model interprets each task, writes terminal commands, reads output, and decides next steps, harvesting AI API keys, SSH credentials, and access tokens along the way.

The botnet scans networks attached to each compromised host every five minutes and propagates to newly found exposed daemons. ThreatDown could not attribute it to a known cluster but points to Costa Rica as a possible operator location. Indicators of compromise include the GH0ST persona file, a CARBONATO_API_KEY environment setting, unexpected Telegram traffic, and reverse SSH tunnels toward AS262145.

The homelab takeaway is direct: keep Docker daemon sockets off untrusted networks (bind to localhost or use TLS with client certificates on 2376), require authentication on any registry, and audit existing hosts for the listed IoCs. Exposed 2375 ports remain one of the most common misconfigurations in home-lab environments.

Source: BleepingComputer.

Roundcube Webmail Under Active Attack via Pre-Auth SQL Injection

Internet-facing Roundcube installations are under active attack exploiting CVE-2026-48842, a pre-authentication SQL injection in the virtuser_query plugin. Canada's Centre for Cyber Security raised the alarm, and security firms have recorded code-injection attacks in the wild — attackers use the flaw to inject malicious code without any valid credentials.

The fix shipped in Roundcube 1.6.16 (LTS branch) and 1.7.1. Any deployment still on an earlier version should be treated as high risk until patched, with log review for injection attempts around the disclosure window. Self-hosted mail stacks — a staple of homelab setups — are disproportionately represented among exposed Roundcube instances, making this one of the more directly relevant items in this week's patch queue.

Sources: The Hacker News and BleepingComputer.

Week in Numbers

• $351.6 million — stolen from Bitget's hot and warm wallets by suspected North Korean hackers; the exchange's ~$464M User Protection Fund is set to cover customer losses.
• Hours, not days — the window between WordPress CVE-2026-87902 disclosure (Sept 23) and first observed exploitation; patched in versions 7.1.2 / 7.0.6 / 6.9.9 / 6.8.10.
• 1,050,000 tokens — GPT-6 Sol's context window, offered at roughly half the list price of its predecessor generation.
• $4/$20 per million tokens — Claude Opus 5.5 API pricing, with cache reads down 60% to $0.20 and output generation over 30% faster than Opus 5.
• Sept 27 / Sept 28 — CISA KEV patch deadlines: WSO2 (CVE-2026-5430) and Adobe Commerce/Magento (CVE-2026-71362) by Sunday; SharePoint (CVE-2026-65660) and Mikrotik RouterOS (CVE-2026-67279) by Monday.