IT News Roundup: PHP2Shell Exploitation, Mistral Large 4 Weights, Binance Heist — Week of Oct 26 – Nov 1, 2026

This week in IT was defined by two parallel narratives: the rapid escalation of the PHP "Black October" vulnerability from patch advisory to confirmed active exploitation, and an unusually dense stretch of open-weight AI releases that pushed frontier-class capability down into hardware homelabbers can actually run. On the security side, CISA added both php2shell CVEs to its Known Exploited Vulnerabilities catalog on Monday, and Palo Alto Networks' Unit 42 confirmed active exploitation against internet-exposed firewall management interfaces — while a use-after-free in the Linux kernel's TCP stack gave researchers (and an unreleased Claude model) a path to unauthenticated remote root. On the AI side, Mistral published open weights for its 1-trillion-parameter Large 4 flagship, Liquid AI shipped the LFM3 family of compact multimodal models with day-zero llama.cpp support, and NVIDIA added Nemotron 3.5 Transcribe to its local-AI lineup. Off the technical front, OpenAI launched GPT-6.1 Terra at half the price of its top-tier model and cut its latest funding target by $10 billion amid a tech selloff driven by AI spending concerns.

Deep Dive: PHP "Black October" Moves From Patch Advisory to Active Exploitation

In late October, two critical vulnerabilities in the PHP core — CVE-2026-109618 and CVE-2026-109619, collectively dubbed "php2shell" or "Black October" by the community — allowed unauthenticated remote code execution on virtually any web server running PHP. Both flaws live in how the engine handles error paths during request processing: per Palo Alto Networks' Unit 42 threat brief (published October 29), CVE-2026-109618 stems from out-of-memory error handling that uses longjmp() and leaves request memory uninitialized, enabling reliable RCE via an OOM trigger; CVE-2026-109619 arises when fatal-error handling in longjmp() leaves global memory pointing at freed regions, so a subsequent request on the same worker can read or branch through them. Both carry a CVSS 4.0 score of 9.2. Patches shipped October 21 for PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35; PHP 8.1 and earlier are end-of-life, with minimal source backports published for 7.4, 8.0, and 8.1.

The blast radius is enormous because PHP underpins a large share of the web's CMS and application stacks. Unit 42 said its Cortex Xpanse telemetry identified more than 22.6 million exposed instances potentially vulnerable as of October 29, and CISA warned that roughly 70% of internet-facing websites run some form of PHP. The Dutch NCSC-NL is tracking the exposure across 929 products — 217 confirmed vulnerable, 165 fixed, 345 under investigation — with its product list updated as recently as October 30.

The story moved from "patch now" to "you may already be compromised" this week. On Monday (October 26), CISA added both CVEs to its Known Exploited Vulnerabilities catalog, obligating federal agencies to remediate within the KEV deadline window. Then on October 29, Unit 42 updated its threat brief with two significant findings: attackers were actively exploiting php2shell against internet-exposed PAN-OS management interfaces, and Dream Research reported a campaign chaining PHP2Shell with a Linux kernel privilege escalation vulnerability to gain root access — tooling that Unit 42 independently verified. SecurityWeek's follow-up noted the attacks successfully rooted Palo Alto firewalls while failing against WordPress deployments, underscoring that appliance and management-plane exposure is where the real damage is landing.

Vendors beyond PHP core have been shipping fixes for embedded PHP stacks throughout the week: Check Point patched its firewall web interface on October 30 (hotfix sk1000302), Cisco released Expressway Select 15.5.21, and Broadcom shipped Carbon Black App Control Server 8.12.2/8.13.2 — all on the same day. AWS also published a security bulletin covering affected services and AMIs.

For IT professionals and homelabbers, the action list is straightforward but urgent: upgrade PHP to the patched releases across every SAPI (mod_php, PHP-FPM, php-cgi, FrankenPHP, LiteSpeed) — mod_php deployments are highest risk because an unauthenticated request can trigger exploitation before application code runs; check vendor advisories for any appliance or management interface with a web UI; and if patching cannot happen immediately, apply the interim mitigations (disable session.upload_progress, restart workers), understanding these reduce but do not eliminate risk. Post-exploitation hunting matters too: Unit 42 published XQL queries to detect post-compromise activity on PAN-OS, and any internet-facing PHP host that was unpatched before October 21 should be treated as potentially compromised — review web access logs for anomalous requests around the patch window and rotate credentials. Sources: Unit 42 threat brief, CISA alert, NCSC-NL php2shell tracker.

AI Models & Weights

Mistral Large 4 Open Weights ("Le Chonk")

Three weeks after its public preview, Mistral published open weights for Large 4 on October 29 — a natively multimodal mixture-of-experts model with roughly 1 trillion total parameters and about 49 billion active per token (the Hugging Face repository is mistralai/Mistral-Large-4-1T-A52B). Trained from scratch in Europe on a cluster of 3,800 NVIDIA Grace Blackwell GPUs, the model supports more than 160 languages natively and ships with three reasoning modes (Low, Medium, High default).

Benchmark results are the strongest open-weight numbers in several categories: 19.7% on FrontierCode 2.0 (top among open weights), 41% on Terminal-Bench 4.0 (+13 points over preview), 51% on Artificial Analysis' Cyber Index, and an 84.5% BrowseComp score; Mistral also reports a 96.6% refusal rate for malicious cyber requests — the highest among open models.

The license is the "Mistral Ecosystem Open-Weight" license, a modified MIT that permits use, modification, fine-tuning, and commercialization without restriction; MaaS providers need a separate commercial agreement. API pricing sits at $0.70/M input ($0.07 cached) and $2.10/M output. Weights are available on Hugging Face with day-one support in vLLM/SGLang, NVIDIA NIM containers, Microsoft Foundry (Azure private deployment), and OpenRouter. Practical note: a 1-trillion-parameter model is datacenter-scale at full precision — the local story will be quantized variants as they land. Source: Mistral announcement.

Liquid AI LFM3 Family

On October 27, Liquid AI released LFM3, a new generation of "Liquid Foundation Models": six ~2B-class open-weight models spanning text, vision, audio, omni (vision + audio + language), and decision tasks on one shared backbone. The architecture is tri-hybrid — gated short convolutions in most layers, sliding-window attention with 256-token windows, and a small number of global-attention layers — pretrained on 22T tokens with the final 3.5T including agentic traces; the text model (LFM3-1.7B) extends context to 256K.

The headline numbers for local use are striking: the 1.73B text model outperforms Liquid AI's previous 2.6B base with roughly 35% fewer parameters, fits in about 1GB at Q4_0 quantization (GGUF), and is the fastest on all three tested devices via llama.cpp — 1.8–2.8x faster than the next-closest model. LFM3-VL-2B matches Gemma 4 12B across an average of 62 vision benchmarks (71.4 vs 71.7) at a fraction of the size, and the audio model transcribes roughly 87x real-time on a single CPU thread.

Everything ships under the LFM Open License v1.0 with day-zero support in llama.cpp, MLX, and vLLM, plus launch partners including AMD, Arm (OpenVINO), Intel, NVIDIA, Qualcomm, and Vercel; LM Studio, Ollama, and Unsloth are part of the builder ecosystem. For homelabbers this is a serious option for always-on local agents: ~1GB RAM at Q4_0 means it runs comfortably on Raspberry Pi-class hardware or as a background service alongside other workloads. Source: Liquid AI blog.

NVIDIA Nemotron 3.5 Transcribe and RTX Spark Local-AI Updates

On October 28, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that handles transcription, speaker attribution, and translation in one promptable model across 40 languages (including seven Indic). With roughly 4.3B active parameters — about 80% fewer than the top-ranked open ASR model — it ranks No. 1 on Hugging Face's Open ASR leaderboard and No. 2 among open weights on Artificial Analysis. Quantized NVFP4 versions are coming via NeMo-Speech.cpp or vLLM for RTX Spark, DGX Spark, Jetson AGX Thor, and Blackwell datacenter GPUs.

The same post covered a batch of local-AI updates relevant to anyone running agents at home: the LFM3 family gets day-zero support on RTX Spark (native llama.cpp/vLLM); NVIDIA App gains resource-management controls that let users allocate GPU compute between foreground apps and background agents (requires hardware-accelerated GPU scheduling in Windows); Nemotron 3.5 Bolt E4B now runs with as little as 4GB of memory across llama.cpp, vLLM, Ollama, Unsloth, and LM Studio; Black Forest Labs' FLUX 3 Image Klein (5.6B open weights) runs on RTX GPUs with as little as 8GB VRAM; and the 64GB DGX Spark is now available from manufacturer partners. Source: NVIDIA blog.

The Rest of the Week

OpenAI Launches GPT-6.1 Terra

On October 26, OpenAI released GPT-6.1 Terra, a new mid-tier model priced at $1/M input and $5/M output ($0.05 cached) — half the price of both GPT-5.6 Terra and the top-tier GPT-6.1 Sol. The positioning is "fast and capable computer use": on Agents' Last Exam v1 it scores 60, matching Astra (Sol: 58.6), OSWorld 2.1 at 53.8, Terminal-Bench 4.0 at 50.3%, with tasks completing roughly 3x faster than Astra or Sol.

The most notable metric is reliability under adversarial prompting: the error rate on prompts designed to induce errors dropped from 53.6% (previous generation) to 16.7%. Terra is available across all paid ChatGPT plans (Plus, Pro, Business, Enterprise, Edu) in ChatGPT Work and Codex, plus via API as gpt-6.1-terra. The launch lands amid a turbulent period for OpenAI's model lineup — GPT-6.1 Astra was reportedly pulled after internal safety testing — making Terra the practical default for agentic workloads at mid-tier pricing. Source: OpenAI announcement.

OpenAI Cuts Funding Target by $10 Billion

On Friday (October 30), the Financial Times reported that OpenAI has cut its latest funding round target by $10 billion, now seeking to raise about $20B at a $1.25T valuation — down from the October 19 ask of $30B at $1.4T, a $150B reset.

The timing is notable: the news landed on a day when the Nasdaq fell 2.9% and AI-linked stocks led a broader tech selloff driven by concerns about the scale of AI capital expenditure. For IT professionals, the signal matters less than the number itself — even at $1.25T, OpenAI remains one of the most valuable private companies in history — but it marks the first visible repricing of frontier-lab valuations amid sustained questions about whether AI revenue growth can keep pace with infrastructure spend. Source: Reuters.

Binance Confirms Record $6.3 Billion Theft

On Thursday (October 29), Binance confirmed that roughly $6.3B in cryptocurrency was stolen from its wallets — the largest crypto heist on record, about four times the $1.5B Bybit hack of February 2025. Per blockchain analysis firm TRM, attackers moved approximately $6.36B across ten blockchains in a burst starting at 20:05 UTC on October 29, most of it within about 20 minutes; Chainalysis independently confirmed the findings with "high confidence."

Binance attributed the breach to a compromised DevOps engineer laptop that gave attackers access to a service container from which they submitted unauthorized withdrawal transactions — no leaked wallet private keys were found. Co-CEO Richard Teng stressed this was "a more complex attack process" rather than a single one-click action, and said additional points of breach had been uncovered. Both TRM and Chainalysis attributed the operation to TraderTraitor, a North Korea state-backed unit also tracked as Lazarus Group/APT38 — the same group behind the Bybit hack.

The response has been a race against stablecoin freezes: Tether, Circle, Ripple, and World Liberty Financial froze about $2.67B in stablecoins, and Arbitrum's Security Council froze 8,426 ETH plus 74.6M ARB — roughly $2.7B total. The attackers raced the freezes by selling USDT to market makers and borrowing ETH against USDT on Aave, leaving them with about $2.4B in assets (420K+ ETH, 18,270 BTC) beyond issuer reach. Binance has suspended deposits and withdrawals while it conducts security checks, revoked authorizations for the affected addresses, and said user funds are safe with losses covered by its SAFU fund; an updated Proof of Reserves is promised once remediation completes. Source: BleepingComputer.

Linux Kernel MPTCP Flaw Enables Remote Root (CVE-2026-98432)

On October 30, BleepingComputer reported that Anthropic's security researchers developed a working exploit for CVE-2026-98432 — a use-after-free in the kernel's IPv4 TCP networking code present since Linux 4.4 (CVSS 9.8 per the kernel CVE team). Unauthenticated attackers can remotely execute commands as root on servers running Multipath TCP listeners; the flaw also affects ordinary TCP connections, though RCE has not been publicly demonstrated without an MPTCP listener.

The technical detail is a race condition in how the kernel handles TCP connection requests: tcp_v4_syn_recv_sock() transfers ownership of ireq->ireq_opt to the child socket without copying it, and another CPU can concurrently process a retransmitted SYN for the same request socket, leaving readers operating on freed memory. Red Hat classifies the issue as potential denial-of-service (kernel crash) with severity 5.9 because exploitation is complex — but Anthropic demonstrated otherwise: an unreleased Claude model developed an exploit that bypassed KASLR and achieved remote root code execution in about 53 minutes against a server running an MPTCP listener.

The exposure is broader than it looks, because Go 1.24+ enables MPTCP by default for TCP listeners on Linux — meaning some Go-based network services may be exposing MPTCP listeners without administrators knowing. Patches are rolling out unevenly as of October 30: Debian fixed it in kernel packages 6.1.190-1 (Bookworm) and 6.12.113-1 (Trixie), while Ubuntu 22.04/24.04/26.04 LTS and RHEL 8/9/10 remain vulnerable with fixes in progress; Oracle Linux, Amazon Linux, and SUSE have shipped partial fixes. Mitigations: restrict access to trusted networks/IPs, and — per CIQ's guidance for Rocky Linux — block IPv4 TCP packets containing IP options at the firewall (nftables examples published) before they reach the vulnerable code path. Disabling MPTCP alone is insufficient because ordinary TCP traffic can still trigger the flaw. No confirmed in-the-wild exploitation has been reported as of publication. Source: BleepingComputer.

Week in Numbers

• $6.36B — amount stolen from Binance wallets on October 29, making it the largest crypto theft on record (~4x the Bybit hack of February 2025)
• 22.6M+ — exposed PHP instances potentially vulnerable to php2shell per Palo Alto Networks' Cortex Xpanse telemetry (as of October 29)
• ~$2.7B — Binance-stolen funds frozen by stablecoin issuers and Arbitrum's Security Council; attackers retain ~$2.4B in unfreezable assets
• $150B — OpenAI's valuation reset in its latest funding round ($30B at $1.4T → $20B at $1.25T)
• 929 products tracked by NCSC-NL for php2shell exposure: 217 vulnerable, 165 fixed, 345 under investigation (as of October 30)