cybr.cx Daily Digest — July 17, 2026

Critical Vulnerabilities

⚠️ Actively exploited — CVE-2026-58644 | Microsoft SharePoint | No CVSS in NVD feed
A deserialization of untrusted data flaw in Microsoft SharePoint allows unauthenticated remote attackers to execute arbitrary code over the network. CISA added this to the KEV catalogue on July 16 with a remediation deadline of July 19 — that's 48 hours away. If your SharePoint deployment isn't patched or isolated, treat this as a P0 right now.

⚠️ Actively exploited — CVE-2026-25089 & CVE-2026-39808 | Fortinet FortiSandbox
Two separate OS command injection vulnerabilities in FortiSandbox, FortiSandbox Cloud, and FortiSandbox PaaS are both confirmed in active exploitation. Both allow unauthenticated attackers to execute arbitrary commands via crafted HTTP requests. The irony of your sandbox being the attack surface shouldn't distract from the urgency: patch or take the management interface off the internet immediately. KEV deadline is July 19.

⚠️ Actively exploited — CVE-2026-56164 | Microsoft SharePoint Server
A missing authentication check for a critical function allows unauthenticated network attackers to elevate privileges on SharePoint Server. The KEV remediation deadline was today, July 17 — if you haven't acted, you're already overdue. Combine the pressure from CVE-2026-58644 and this one: SharePoint is under concentrated, active attack right now.

⚠️ Actively exploited — CVE-2026-46817 | Oracle E-Business Suite
An improper privilege management flaw in Oracle E-Business Suite's Payments module allows unauthenticated HTTP access to result in full takeover of Oracle Payments. CISA deadline is July 18. Organisations running Oracle EBS for financial processing should consider this critical regardless of CVSS framing — payment system compromise is a direct path to fraud.

⚠️ Actively exploited — CVE-2026-15409 & CVE-2026-15410 | SonicWall SMA1000
An SSRF flaw (CVE-2026-15409) and a code injection flaw (CVE-2026-15410) are both actively exploited in SonicWall SMA1000 appliances. The code injection requires authenticated admin access but the SSRF is unauthenticated, creating a plausible chained attack path. CISA deadline was July 17. SonicWall edge devices have been a persistent target — patch or remove from external exposure now.

⚠️ Actively exploited — CVE-2026-56155 | Microsoft Active Directory Federation Services
An insufficient access control flaw in AD FS allows a locally authenticated attacker to elevate privileges. Less dramatic than the network-remote SharePoint flaws but critical in post-breach scenarios — an attacker with any local foothold can weaponise this for privilege escalation toward domain-level access.

⚠️ Actively exploited — CVE-2023-4346 | KNX Protocol (Building Automation)
A lockout mechanism flaw in KNX Protocol Connection Authorization Option 1 allows an attacker to purge all devices and set a BCU key, effectively bricking or locking building automation systems. This is an OT/ICS issue — affects HVAC, lighting, access control infrastructure in commercial buildings. If you manage smart building environments without additional KNX security layers, audit exposure now.

CVE-2026-63085 | Axelor Open Platform | CVSS 8.8
Authenticated non-admin users in Axelor versions prior to 8.2.2 can escalate privileges by exploiting unenforced field restrictions on nested relational save operations — effectively writing to User record fields like roles and group via a related entity's save path. This bypasses the USER_REST restriction layer entirely. Organisations using Axelor for ERP or business process management should treat any authenticated user as a potential admin until patched.

CVE-2026-15008 | WordPress — Uncanny Automator Plugin (≤ 7.3.1.4) | CVSS 8.1
Unauthenticated attackers can delete arbitrary files on the server through insufficient path validation in the fr_token function. Deleting configuration files or core WordPress files creates a straightforward path to remote code execution via reinstallation tricks. With WordPress powering a significant chunk of the web, widely-installed automation plugins are high-value targets — update immediately.

CVE-2026-63305 | AVideo (≤ 29.0) | CVSS 8.1
The ffmpeg.json.php endpoint concatenates user-supplied notifyCode and callback parameters directly into shell commands without escaping. An attacker who can craft a valid encrypted payload gets OS command execution as the web server user. Self-hosted video platforms are frequently deprioritised in patch cycles — check your AVideo version now.

CVE-2026-12753 | WordPress — Advance Product Search for WooCommerce (≤ 1.4.4) | CVSS 7.5
Unauthenticated SQL injection via the s and match parameters due to insufficient escaping. Attackers can read or manipulate the entire database without credentials. Any WooCommerce store running this plugin should update or disable it pending a patch.

Headline News

Teenagers Who Live-Streamed the TfL Hack Are Jailed — and Still Hacking From Prison

Two British teenagers have been jailed for their roles in a cyberattack on Transport for London that caused an estimated £29 million in damage — one of the most significant infrastructure attacks in UK history. The pair, described as having amassed cryptocurrency wealth through prior hacking activity, conducted and reportedly live-streamed elements of the intrusion from their bedrooms. What makes the case operationally notable is the attack's reach: TfL was forced to take systems offline, affecting contactless payment data for roughly 5,000 customers and exposing some bank account details through the Oyster online service. More striking still is a detail emerging from coverage: at least one of the individuals was allegedly continuing hacking activity from inside custody, raising serious questions about device access policies in juvenile detention facilities. For practitioners, this case underscores that capable threat actors are not always nation-state actors with sophisticated infrastructure — motivated teenagers with OPSEC gaps (one was reportedly identified partly through a food delivery order) can inflict enterprise-scale damage. The social engineering and access techniques used highlight the continued effectiveness of low-tech human factors against high-value targets.

AI Coding Assistants Can Be Turned Into Botnet Infrastructure via Hallucination Exploitation

Researchers have disclosed a novel attack class demonstrating that popular AI development tools — including widely-used code generation and completion platforms — can be manipulated into creating botnet infrastructure by exploiting model hallucinations. The technique involves crafting prompts that cause the AI to generate code referencing non-existent but plausible package names, which attackers can pre-register with malicious payloads — a variant of dependency confusion scaled through AI trust. Nine major AI platforms were found susceptible in the research. For security teams overseeing developer environments, this significantly expands the supply chain threat model: it's no longer sufficient to vet dependencies a human developer explicitly chose, because AI-generated code may silently introduce attacker-controlled packages the developer never consciously selected. The finding adds urgency to calls for AI output sandboxing, mandatory dependency review gates in CI/CD pipelines, and awareness training for developers using AI pair-programming tools. This is an emerging attack surface that most organisations have no current control for.

Schrödinger's Feed

Physicists are revisiting one of quantum mechanics' foundational assumptions — that complex (imaginary) numbers are a mathematical necessity rather than a convenience. A new review suggests that in atomically thin quantum materials, where light-generated excitons interact directly with magnetic behaviour, quantum states can be described and manipulated in ways that challenge previous theoretical frameworks. The practical upshot for cryptography isn't immediate, but foundational shifts in how we model quantum systems have historically preceded shifts in what quantum computers can efficiently compute. Practitioners building long-term PQC migration roadmaps should stay alert: if the mathematical underpinnings of quantum advantage are still being refined, so too are the threat timelines we're planning against.

/dev/random

Someone has published a detailed walkthrough on training a generative AI kick drum synthesiser model on a six-year-old Linux desktop with a consumer GPU carrying 6GB of VRAM — the kind of hardware that was considered marginal for gaming, let alone machine learning. The process involves quantisation tricks, gradient checkpointing, and a custom diffusion architecture tuned specifically for the narrow frequency profile of a kick drum. It works. The security angle, if you squint: the same "run capable ML models on bargain hardware" techniques being refined for drum machines are directly applicable to running local LLMs for offensive tooling, with no API logging, no rate limits, and no content filters. Your threat model just got a basement upgrade.