Laravel auth bypass leads a week of critical web flaws
The Laravel auth bypass is among the most urgent web security issues disclosed this week. CVE-2026-114509, rated 9.2 on the 10-point Common Vulnerability Scoring System (CVSS), affects the framework’s built-in TokenGuard authentication component. Under some configurations of two related open-source database servers, a specially formed credential that is not ordinary text can cause an authentication query to select an unintended user instead of rejecting the request. An unauthenticated attacker could consequently enter an application as an existing user without knowing that person’s application programming interface (API) token. Laravel fixed the issue in versions 13.29.0 and 12.68.0. Website operators using TokenGuard should treat the updates as account-takeover fixes rather than routine maintenance.
Other internet-facing applications also received serious vulnerability disclosures. Mobeta reported that an unauthenticated SQL injection flaw in GLPI 11.0.8, tracked as GHSA-753w-2942-4pxm, can expose the full application database. SQL injection means malicious input is interpreted as a database command. Depending on the affected version, the access could progress to administrator takeover and remote code execution (RCE), which allows commands to run on the server. GLPI 11.0.9 contains the patch. Rapid7 separately found three Eptura Archibus flaws: CVE-2026-107186 bypasses authentication through a legacy API, CVE-2026-107187 permits full database reading, and CVE-2026-107188 provides an unsafe file upload that can lead to RCE as NT AUTHORITY/Local Service.
AegisSec Research disclosed CVE-2026-65162, also called ShadowForge, in NVIDIA AIStore for Linux. The CVSS 10.0 flaw is present in a legacy ETL Kubernetes PodSpec initialization API. Kubernetes manages software packaged into isolated containers, and the researchers said this weakness permits creation of a privileged container followed by compromise of the underlying host. Developer systems face another route to code execution through OpenCode on Windows and Windows Subsystem for Linux. Pillar Security found that an attacker-controlled remote Model Context Protocol server could execute code during remote OAuth authentication without another approval step. Anomaly fixed that flaw in OpenCode 1.18.33. The disclosure followed an OpenCode Console breach linked to a large-scale distillation attack that affected 1.4% of Console users and potentially exposed API keys associated with Anthropic, OpenAI and Google AI models.
Roundcube released versions 1.6.20 and 1.7.5 to correct 20 security problems. These include IMAP command injection, in which crafted input reaches the mail protocol as an instruction, a server-side request forgery (SSRF) filter bypass that can make a server contact unintended destinations, and unsafe handling of attachments. No exploitation of those patched Roundcube vulnerabilities has been reported. Chrome also expanded a transport security measure to all users by default: the browser now asks permission before connecting through unencrypted HTTP. That warning had previously applied to people using Enhanced Protection in Safe Browsing.
A separate critical flaw affects Signal calling across Android, iOS and desktop. CVE-2026-113380 has a CVSS score of 9.3 and resides in RingRTC, the WebRTC-based library used for voice and video calls. According to the researcher, a malicious caller can flood the call’s event queue so that the victim’s hang-up request remains stuck behind attacker-generated events. Audio, and potentially video, may therefore remain active after the victim tries to end the call. The reported weakness shows why closing an interface element is not always equivalent to stopping the underlying process.
Credential theft and exposed secrets produced several significant incidents. Zeabur said attackers used leaked internal credentials stored as plain text to enter a shared cluster, some Dedicated Servers and internal systems on October 14 and 15, 2026. The attackers reached the platform database and stole more than 3.2 million environment-variable records plus information for 321,810 accounts. The development platform plans to begin shutting down on December 1 and complete the process by December 31, 2026. It was Zeabur’s second breach of the year, following an August incident involving account information, API keys and other secrets.
PingCAP reported that stolen service-account credentials enabled access to internal systems between September 9 and October 7, 2026. The intruders copied source code and records containing names, email addresses, organization and project relationships, SQL editor text and errors, customer contacts, contract and order details, pricing, payment terms, billing configuration and encrypted API-key values. PingCAP discovered the activity and began disabling affected credentials on October 7. Andromeda identified unauthorized credential access on October 13 after activity it believes began around October 9. It disabled the compromised accounts, notified customers it considered affected and said it found no evidence that customer models, datasets or training data were accessed. Genesys also disclosed unauthorized access to internal customer relationship management and learning systems. Business and account information was accessed or downloaded, but the company said its Genesys Cloud, Genesys Engage and Genesys Engage Cloud platforms were not compromised.
Phishing remains an effective route around otherwise strong account controls. Bluesky said more than 4,500 accounts were compromised through posts advertising a fake government payment; it temporarily suspended affected accounts while working to restore access. Dataminr identified fake Okta and ServiceNow portals at riskified-okta[.]com and shipmonk-servicenow[.]com targeting e-commerce software providers. The pages attempted an MFA downgrade, meaning users of hardware-based multi-factor authentication (MFA) were pushed toward weaker fallback methods such as text-message or email codes where account policy allowed them. Singapore Police also reported at least 62 Android malware scam cases since April 2026, with losses of S$989,000. Victims were directed from social media advertisements to install an Android Package Kit (APK) file before callers impersonating officials or bank staff sought money, banking details, personal information or valuables.
AI-connected development access creates additional risks when permissions are broad. A researcher demonstrated that hidden instructions in an attacker-controlled public repository could manipulate ChatGPT’s GitHub Connector during a deployment review. With the default all-repositories permission, the connector could write a GitHub Actions workflow into a separate private repository, where automated code could access and transmit stored secrets. The attack required neither access to the victim’s repository nor a malicious uploaded document.