cybr.cx Daily Digest — October 23, 2026

Critical Vulnerabilities

⚠️ Actively exploited — CVE-2026-106931 | EG Secure Solutions SiteGuard Server Edition | CVSS: N/A
An OS command injection flaw in SiteGuard Server Edition allows unauthenticated attackers to execute arbitrary commands with root or SYSTEM privileges. CISA's remediation deadline was today — if you haven't patched, you're already overdue. Treat any internet-facing SiteGuard instance as potentially compromised.

⚠️ Actively exploited — CVE-2026-96652, CVE-2026-96654, CVE-2026-96655, CVE-2026-96656 | Plex Media Server | CVSS: N/A
Four Plex vulnerabilities are being actively exploited in a chain: an SSRF via the protocol parameter (CVE-2026-96652) that works with any valid X-Plex-Token, a second SSRF via the /video/:/transcode/universal path (CVE-2026-96655), a URI scheme injection enabling cross-plugin function calls (CVE-2026-96654), and a path traversal enabling arbitrary file writes that execute on load (CVE-2026-96656). Chained together, these allow an attacker to pivot from a low-privilege token to remote code execution. Plex deployments exposed to internal networks are at particular risk of lateral movement.

⚠️ Actively exploited — CVE-2026-102764 | BeyondTrust Remote Support | CVSS: N/A
An authentication bypass via spoofing in BeyondTrust Remote Support allows unauthenticated remote attackers to perform file operations on the appliance. CISA's due date passed on October 19 — escalate patching immediately, as privileged access management tools are high-value targets.

⚠️ Actively exploited — CVE-2025-61624 | Fortinet FortiOS, FortiPAM, FortiProxy, FortiSwitchManager | CVSS: N/A
A path traversal flaw allows authenticated admins with read-write permissions to write or delete arbitrary files via specific CLI commands. In environments where admin credentials may already be compromised or shared, this represents a significant persistence and destructive capability.

⚠️ Actively exploited — CVE-2026-21589 | Multiple Atlassian Data Center Products | CVSS: N/A
Unauthenticated external parties can access files or directories across Bitbucket, Confluence, Jira Software, Jira Service Management, Bamboo, Crowd, Crucible, and Fisheye Data Center deployments. The breadth of affected products makes this critical for any enterprise running Atlassian's on-premise stack.

⚠️ Actively exploited — CVE-2026-110936 | Siemens Software Center | CVSS: N/A
An uncontrolled search path element vulnerability allows authenticated local attackers to plant a malicious DLL that executes with application-level privileges. Particularly concerning in OT/ICS environments where Siemens software is common and local access is not always tightly controlled.

CVE-2026-76291 | Splunk Enterprise (below 10.4.4 / 10.2.8 / 10.0.11 / 9.4.16) | CVSS: 8.8 — HIGH
Users with the schedule_search capability can bypass workflow safeguards and run arbitrary risky SPL commands, enabling data disclosure, modification, or denial of availability. In many Splunk deployments, scheduled search permissions are broadly granted — audit your role assignments now.

CVE-2026-114658 | Grav Login Plugin (3.6.0–3.9.12) | CVSS: 8.8 — HIGH
Page authors can craft an invitation form in page frontmatter that grants super-admin access, then submit it unauthenticated, handing full site control to an arbitrary email address. Any Grav CMS instance with the Login plugin below 3.9.13 and user-editable pages should be treated as at risk of full compromise.

CVE-2026-114635 / CVE-2026-114636 | Traefik (before v2.11.58 / v3.7.14) | CVSS: 8.7 — HIGH
Two separate authentication bypass issues: the first allows unauthenticated HTTP/2 clients on Go 1.27 builds to skip mTLS enforcement entirely when TLS state is nil; the second bypasses access-control annotations on Kubernetes Ingress NGINX ssl-passthrough routes. Either flaw can expose backend services that operators believe are protected. Upgrade to v2.11.58 or v3.7.14.

CVE-2026-115328–115332 | JSON-SCADA (<0.65-alpha) and Lotus (≤0.10.2) | CVSS: 8.8 — HIGH
Multiple SQL injection flaws across two platforms. JSON-SCADA's opcApi() function accepts unsanitised NodeId.Id, StartTime, and EndTime values, enabling stacked SQL statements against historian data. Lotus has three separate injection paths — metric queries, billing property names, and subscription filter keys — all rendered unescaped into database queries shared across tenants. Multi-tenant data isolation is effectively nonexistent in unpatched versions.

Headline News

AI-Assisted Discovery Uncovers Critical PHP Vulnerability Chain

A coordinated vulnerability research effort, significantly aided by large language model tooling from multiple AI providers, has surfaced a set of critical flaws in PHP affecting a substantial portion of internet-facing web infrastructure. The vulnerabilities were serious enough that Amazon's threat intelligence team independently observed exploitation activity — dubbed PHP2Shell — in the wild prior to coordinated public disclosure, catching it through their MadPot sensor network. What makes this incident notable for practitioners is the dual-edged nature of the AI involvement: the same models accelerating defensive research are equally available to threat actors hunting for exploitable patterns at scale. Defenders should audit PHP version deployments immediately and apply available patches; the fact that in-the-wild exploitation preceded public disclosure suggests the window between discovery and weaponisation is compressing further.

Railway Platform Security Incident Exposes Infrastructure Risk

Cloud infrastructure platform Railway disclosed a security incident on October 21st, with details indicating unauthorised access to internal systems. While full scope details are still emerging, the disclosure is significant because Railway is widely used by developers to host production workloads, meaning any compromise of platform-level infrastructure carries supply-chain implications for downstream applications. The incident follows a pattern of developer platform targeting — attackers increasingly recognise that compromising the tooling layer yields access to many environments simultaneously. Practitioners running workloads on Railway should review their deployment secrets, environment variables, and any credentials that may have transited the platform.

1.5 Million Crime Stoppers Tips Leaked, Exposing Confidential Informants

Over 1.5 million tips submitted to Canadian Crime Stoppers — a programme explicitly designed to guarantee anonymity to informants — have been exposed in a data breach, according to reporting that includes verification with individual tipsters. At least some records contain submitters' full names, directly contradicting the programme's core anonymity guarantees and potentially placing informants at physical risk. The breach is a stark reminder that privacy-by-design must be validated technically, not just promised in policy: systems handling sensitive informant data require rigorous access controls, data minimisation, and regular third-party audits. For security practitioners working with law enforcement or government clients, this case should be referenced as a concrete example of the real-world human consequences of inadequate data governance.

Schrödinger's Feed

IBM and Rutgers University have demonstrated real-time quantum feedback control across 100 qubits on IBM's Heron processor, with findings published in Nature Physics. The experiment implemented adaptive quantum circuits to validate non-unitary dynamics — a technical milestone directly relevant to scalable quantum error correction, which is the primary engineering barrier between today's noisy processors and cryptographically relevant machines. Error correction is the unlock: without it, quantum computers cannot sustain the coherence needed to run Shor's algorithm at the scale required to threaten current public-key cryptography. Practitioners invested in PQC migration timelines should watch this space — each credible advance in error correction shortens the theoretical runway before cryptographically relevant quantum computing becomes plausible.

/dev/random

NVIDIA has open-sourced something called OpenShell — billed as a "safe, private runtime for fleets of autonomous AI agents." The security community's collective eyebrow raise is understandable: a sandboxed shell environment for AI agents that can, presumably, run commands, manage files, and coordinate across a fleet, is exactly the kind of thing that security researchers will spend the next six months trying to escape from. The project essentially formalises the threat model that red teamers have been sketching on whiteboards for two years — autonomous agents need a runtime, runtimes have boundaries, and boundaries are a challenge. Watch for the first "OpenShell escape" CVE with the same energy you'd watch a newly launched bug bounty programme.