Blahalla

ErrTraffic Exploits Visual Page Breaks to Fuel ClickFix Attacks, Rebranding Exploits as “GlitchFix”

gbhackers.com · 6 months ago

A new social engineering technique called ErrTraffic is being used to trick users into installing malicious software. This technique, nicknamed "GlitchFix," intentionally breaks website visuals to simulate a device malfunction, prompting victims to run fake updates. The tool, sold on Russian forums for around $800, boasts a remarkably high success rate, with nearly 60% of users clicking on malicious buttons. ErrTraffic delivers legitimate Remote Monitoring and Management tools like FleetDeck and ITarian, bypassing traditional security defenses.

Trump team touts a coming economic revolution as voters fear job losses

www.washingtonpost.com · 6 months ago

false

Pro-Russian denial-of-service attacks target UK, NCSC warns

www.bitdefender.com · 6 months ago

The UK's National Cyber Security Centre has issued a warning about distributed denial-of-service attacks targeting British organizations. These attacks, linked to pro-Russian hacking groups, are disrupting online services, particularly those related to local government and critical infrastructure. While typically not highly sophisticated, these attacks can cause significant disruption and cost. The NCSC advises organizations to review defenses and harden resilience, especially those using operational technology, and mentions groups including Cyber Army of Russia Reborn, Z-Pentest, NoName057(16), and Sector16.

Oracle’s First 2026 CPU Delivers 337 New Security Patches

www.securityweek.com · 6 months ago

Oracle released its first Critical Patch Update for 2026, addressing 337 security patches across more than 30 products. The update includes fixes for approximately 230 unique CVEs, with more than two dozen resolving critical-severity vulnerabilities that can be exploited remotely. Several patches target CVE-2025-66516, a critical Apache Tika defect allowing XML injection attacks, impacting products like Commerce and Fusion Middleware. Oracle Communications received the most fixes, totaling 56, while Fusion Middleware followed with 51.

Critical Zoom Vulnerability Enables Remote Code Execution via Command Injection

gbhackers.com · 6 months ago

A critical command injection vulnerability, identified as CVE-2026-22844, has been discovered in Zoom Node Multimedia Routers, potentially allowing remote code execution. This flaw affects Zoom Node Meetings Hybrid and Meeting Connector deployments running module versions prior to 5.2.1716.0. The vulnerability has a CVSS score of 9.9, indicating severe exploitability and minimal attack complexity. Zoom recommends immediately updating to version 5.2.1716.0 or later to mitigate the risk of unauthorized code execution.

Vulnerability prioritization beyond the CVSS number

www.csoonline.com · 6 months ago

Traditional vulnerability assessments rely heavily on the Common Vulnerability Scoring System, or CVSS, but this approach often overlooks crucial context. Events like the Equifax breach, SolarWinds attack, and Log4Shell highlight that vulnerabilities can spread through interconnected systems, rendering high CVSS scores misleading. A new model, the unified linkage model, focuses on relationships between systems, adjacency, inheritance, and trust, to better understand propagation pathways and prioritize vulnerabilities based on their potential impact, rather than just severity. Log4Shell became significant not due to its technical severity, but because of its widespread inheritance and dependency across systems.

Exposure Assessment Platforms Signal a Shift in Focus

thehackernews.com · 6 months ago

Traditional vulnerability management is proving ineffective for modern enterprises, prompting the emergence of Exposure Assessment Platforms, or EAPs. These platforms shift focus from simply identifying vulnerabilities to managing continuous threat exposure by tracking how systems, identities, and vulnerabilities interact. Gartner's new Magic Quadrant for EAPs evaluates vendors based on their ability to prioritize risk and provide integrated visibility across environments, aiming to reduce unplanned downtime by 30% by 2027. EAPs consolidate discovery, prioritize based on context, integrate exposure data into workflows, and support lifecycle tracking, moving beyond static vulnerability lists to model attacker behavior.

GNU InetUtils Vulnerability Exploited via “-f root” to Achieve Full System Control

gbhackers.com · 6 months ago

A critical vulnerability exists in versions 1.9.3 through 2.7 of GNU InetUtils’ telnetd server, allowing attackers to gain root access without credentials. The flaw, introduced in 2015, arises from improper sanitization of the USER environment variable, enabling exploitation via a specially crafted string containing “-f root.” Attackers can bypass authentication and directly obtain a root shell with a simple telnet command. Patches addressing the vulnerability are available, and disabling telnetd entirely in favor of SSH is recommended.

Magecart Hack Injects JavaScript to Steal Online Payment Data

gbhackers.com · 6 months ago

A new Magecart-style attack is targeting e-commerce websites through malicious JavaScript injections. The script, initially hosted at cc-analytics[.]com/app.js, steals payment card data during checkout by intercepting information entered into checkout fields and transmitting it to attacker-controlled servers. Researchers discovered a broad attack infrastructure across more than 30 domains, using techniques like URLScan and WHOIS queries to map the attacker’s network. Organizations are advised to monitor for unexpected script tags, suspicious network requests, and implement security measures like content security policies to protect against this threat.

Analysis of 6 Billion Passwords Shows Stagnant User Behavior

www.securityweek.com · 6 months ago

An analysis of six billion compromised passwords reveals that simple numeric sequences and common words like "123456" and "password" remain popular choices, demonstrating little improvement in user behavior. The most common passwords also include default credentials frequently used on networking equipment and IoT devices, potentially granting attackers access to critical systems. Malware such as Lumma and RedLine were frequently used to steal these credentials. Experts recommend continuous monitoring, blocking predictable patterns, and implementing strong authentication methods to mitigate this ongoing risk.

Chainlit AI Framework Flaws Enable Data Theft via File Read and SSRF Bugs

thehackernews.com · 6 months ago

Security researchers have identified critical vulnerabilities in the Chainlit AI framework, potentially enabling attackers to steal sensitive data and escalate privileges. These flaws, CVE-2026-22218 and CVE-2026-22219, involve an arbitrary file read and a server-side request forgery within the project element update flow. Attackers could exploit these vulnerabilities to access API keys, credentials, and even source code, allowing for lateral movement within an organization. Separately, a vulnerability dubbed MCP fURI has been found in Microsoft’s MarkItDown server, enabling unauthorized access to resources and potentially exposing AWS account credentials.

EU-Kommission will Huawei und ZTE aus Netzen verbannen

www.csoonline.com · 6 months ago

The European Commission is proposing new measures to enhance cybersecurity across the EU, potentially banning network equipment from companies like Huawei and ZTE. This initiative aims to address concerns about espionage and sabotage by foreign entities and follows prior recommendations to limit the use of these vendors' technology in 5G networks. The proposal would allow the Commission to assess risks associated with specific manufacturers and prohibit their equipment in critical infrastructure, requiring replacements within three years. Additionally, the EU intends to strengthen the European Union Agency for Cybersecurity, ENISA, to combat ransomware attacks and establish cybersecurity standards.

VoidLink Linux Malware Framework Built with AI Assistance Reaches 88,000 Lines of Code

thehackernews.com · 6 months ago

Researchers have discovered a sophisticated Linux malware framework called VoidLink, developed with significant assistance from artificial intelligence. The framework, written in Zig, reached over 88,000 lines of code and is believed to have been created by a single individual with kernel development expertise. Analysis revealed the developer likely used a large language model to accelerate development, generating code and documentation based on specified plans. This demonstrates how AI is lowering the barrier to entry for creating advanced malware, enabling even individual actors to rapidly build complex systems.

PURELOGS Payload Hidden in Weaponized PNG Images Used in Stealth Attacks

gbhackers.com · 6 months ago

A new PURELOGS infostealer campaign is using weaponized PNG image files to evade detection. Attackers use a JScript file, disguised as a pharmaceutical invoice, to launch a hidden PowerShell process which downloads a seemingly harmless PNG image from archive.org. The PNG image contains a Base64-encoded payload hidden between custom markers, which is then executed in memory to steal credentials from Chromium-based browsers, cryptocurrency wallets, and other applications. This technique utilizes fileless execution and legitimate infrastructure to bypass traditional security controls.

Weaponized Shipping Documents Spread Remcos RAT in Stealthy Malware Campaign

gbhackers.com · 6 months ago

A sophisticated phishing campaign is using fake shipping documents to distribute a fileless variant of the Remcos RAT. The emails trick recipients into opening malicious Word documents that download and execute a remote template, exploiting CVE-2017-11882, a Remote Code Execution vulnerability in Microsoft’s Equation Editor. The campaign uses multiple URL shorteners to conceal the malicious RTF file and ultimately injects the Remcos agent into memory, avoiding file-based detection. Remcos offers extensive capabilities including system resource management, surveillance, and network control.

Fake PNB MetLife payment pages abusing UPI & Telegram bots

malwr-analysis.com · 6 months ago

Scammers are impersonating PNB MetLife Insurance to steal customer details and redirect payments through fake payment gateways. The initial pages request policy numbers and mobile numbers, which are then sent to Telegram bots. Victims are subsequently directed to enter payment amounts and complete transactions via QR codes or deep links within legitimate UPI apps. A more advanced phishing template expands the scope to include full banking and card data harvesting, also sent to Telegram. Attackers are leveraging free hosting platforms like EdgeOne Pages to rapidly deploy and modify these phishing sites.

13 cyber questions to better vet IT vendors and reduce third-party risk

www.csoonline.com · 6 months ago

There are 13 questions security leaders should ask IT vendors to reduce third-party risk, covering attestations such as SOC 2 Type II and ISO 27001, how controls are maintained and materially changed, who can alter identity posture and how social engineering is prevented, and evidence of onboarding, offboarding, and access reset workflows. Providers should also disclose independent testing and its frequency, provide a complete inventory and lifecycle management for OAuth integrations and privileged APIs, define contractual and operational commitments if processes are abused, and supply session controls, monitoring, and isolation mechanisms for customer assets. The recommendations follow high-profile supply-chain incidents, including a helpdesk-related breach that led a major retailer to end an outsourcing deal, OAuth token compromises that exposed credentials and data from as many as 700 organizations, and a critical zero-day in an enterprise resource platform that underscored vendor-targeted risk. Security leaders are urged to require ongoing, evidence-based verification, patching SLAs, the ability to test vendor processes, rapid breach notification, and sufficient cyber insurance to cover multi-customer incidents.

oss-sec: GNU InetUtils Security Advisory: remote authentication by-pass in telnetd

seclists.org · 6 months ago

A security vulnerability exists in GNU InetUtils telnetd, versions 1.9.3 through 2.7. The telnetd server passes user-supplied environment variables directly to the login process, potentially allowing an attacker to bypass authentication by crafting a "USER" environment variable containing "-f root". This results in automatic login as the root user. It is recommended to disable the telnetd service, restrict network access, or apply a patch to sanitize variable expansion.

LastPass Warns of Fake Maintenance Messages Targeting Users’ Master Passwords

thehackernews.com · 6 months ago

LastPass is warning users about an active phishing campaign impersonating the password management service. Attackers are sending emails claiming upcoming maintenance and requesting users to back up their vaults within 24 hours. These emails direct victims to a malicious website and LastPass emphasizes that they will never request a user’s master password. The emails originate from several suspicious addresses including support at sr22vegas dot com, support at lastpass dot server8, support at lastpass dot server7, and support at lastpass dot server3.

For cyber risk assessments, frequency is essential

www.csoonline.com · 6 months ago

false

Hackers Exploit Visual Studio Code to Deploy Malicious Payloads on Victim Systems

gbhackers.com · 6 months ago

Hackers are exploiting Microsoft Visual Studio Code to deploy malicious payloads on victim systems, as part of the Contagious Interview campaign. The attack begins when victims open malicious Git repositories, which then prompts Visual Studio Code to automatically process configuration files containing embedded commands. These commands retrieve and execute JavaScript payloads from remote servers, enabling remote code execution, system fingerprinting, and persistent command-and-control communication. Threat actors are using vercel.app to host payloads and employing obfuscation techniques to evade detection.

When The Gateway Becomes The Doorway: Pre-Auth RCE in API Management

principlebreach.com · 6 months ago

A critical vulnerability has been discovered in API management systems, allowing attackers to remotely execute code. The flaw occurs during the pre-authentication phase, bypassing security checks. This allows attackers to gain unauthorized access and control over the system. The vulnerability is identified as CVE-2024-3093.

CERT/CC Warns binary-parser Bug Allows Node.js Privilege-Level Code Execution

thehackernews.com · 6 months ago

A vulnerability has been identified in the binary-parser npm library, allowing for the potential execution of arbitrary JavaScript code. The flaw, CVE-2026-1245, stems from insufficient validation of user-supplied input during dynamic code generation. This could enable attackers to execute code with Node.js process privileges, potentially accessing local data or manipulating application logic. Users are advised to update to version 2.3.0 to resolve the issue.

Chrome 144 Released to Fix High-Severity V8 JavaScript Engine Flaw

gbhackers.com · 6 months ago

Researchers have discovered three methods attackers can exploit to bypass Apple's Lockdown Mode. These techniques involve manipulating the system clipboard, exploiting vulnerabilities in Universal Control, and leveraging certain features of HomeKit. The clipboard manipulation allows attackers to potentially execute malicious code. Universal Control's file sharing functionality can be exploited, and vulnerabilities in HomeKit could be used for unauthorized access. CVE-2023-42079, CVE-2023-42080, and CVE-2023-42081 are associated with these findings.

Azure Private Endpoint Deployments Expose Cloud Resources to DoS Attacks

gbhackers.com · 6 months ago

A critical architectural flaw in Azure’s Private Endpoint deployments can lead to denial of service attacks against cloud resources. The vulnerability occurs when Private DNS zone resolution interacts poorly with hybrid networking, potentially impacting Azure storage accounts and other critical services. Attackers, or even unintentional configurations by administrators or vendors, can trigger this issue by creating Private Endpoints that disrupt DNS resolution. Microsoft suggests mitigation strategies including enabling fallback DNS resolution and manually adding DNS records, but acknowledges the issue as a known limitation.

Everest Ransomware Group Allegedly Claims Breach of McDonald’s India Systems

gbhackers.com · 6 months ago

Researchers have identified a critical vulnerability in the widely used Atlassian Confluence collaboration platform. The flaw, specifically a server-side request forgery, allows unauthenticated attackers to download Confluence backups. This issue, designated as CVE-2023-39602, impacts versions prior to 7.19.16, 8.3.4, and 8.4.4. Atlassian has released patches and workaround instructions to address the security risk.

Threat Actors Exploit LinkedIn for RAT Delivery in Enterprise Networks

gbhackers.com · 6 months ago

Cybercriminals are exploiting LinkedIn private messages to deliver remote access trojans to enterprise networks. Attackers use weaponized open-source tools, including a PDF reader, portable Python interpreter, and WinRAR archives, to bypass security detections through techniques like DLL sideloading and in-memory decryption. The campaign leverages LinkedIn’s professional context for targeted attacks on high-value individuals and utilizes Base64-encoded shellcode runners to avoid disk-based artifacts. Security teams should implement social media security awareness training and application control policies to mitigate this evolving threat.

Three vulnerabilities in Anthropic Git MCP Server could let attackers tamper with LLMs

www.csoonline.com · 6 months ago

Researchers have identified three vulnerabilities in Anthropic’s official Git MCP server that could allow attackers to manipulate large language models. These vulnerabilities, CVE-2025-68143, CVE-2025-68145, and CVE-2025-68144, could enable unauthorized code execution and data tampering. The risk is amplified when both the Git MCP server and Filesystem MCP server are enabled, creating a critical vulnerability. Developers are urged to update to version 2025-12.18 and monitor systems for unexpected.git directories.

OpenAI's ChatGPT Atlas browser is testing actions feature

www.bleepingcomputer.com · 6 months ago

false

Flaws in Chainlit AI dev framework expose servers to compromise

www.csoonline.com · 6 months ago

Security researchers have identified two vulnerabilities in the Chainlit AI development framework, potentially allowing attackers to compromise servers. The flaws, CVE-2026-22218 and CVE-2026-22219, enable attackers to read arbitrary files and trigger cross-site request forgery attacks. Exploitation can lead to the exposure of sensitive data like API keys, database content, and potentially cloud credentials, facilitating lateral movement within an organization’s infrastructure. Patches were released in version 2.9.4, and detection signatures are available for security tools.

Google says Gemini won’t have ads, as ChatGPT prepares to add them

www.bleepingcomputer.com · 6 months ago

false

Tech workers ask their bosses to ‘call the White House’ over ICE raids

www.washingtonpost.com · 6 months ago

Technology workers are urging company leaders to intervene with the White House regarding increased immigration enforcement actions. A public letter, signed by hundreds of employees, requests that CEOs contact the administration to limit the deployment of federal immigration agents. The letter was published Tuesday and includes signatures from staff at several major technology companies. The workers believe corporate leaders have a responsibility to advocate for policy changes.

OpenAI rolls out age prediction model on ChatGPT to detect your age

www.bleepingcomputer.com · 6 months ago

false

ACF plugin bug gives hackers admin on 50,000 WordPress sites

www.bleepingcomputer.com · 6 months ago

A critical vulnerability exists in the Advanced Custom Fields Extended plugin for WordPress, potentially allowing unauthenticated attackers to gain administrative access. The flaw, identified as CVE-2025-14533, arises from insufficient role restrictions when creating or updating users through forms. Approximately 50,000 WordPress sites are believed to be exposed, although the vulnerability requires specific configurations to be exploitable. Threat monitoring firm GreyNoise has observed widespread reconnaissance activity targeting WordPress plugins, including the Advanced Custom Fields Extended plugin.

Snap Settles Lawsuit on Social Media Addiction, Avoiding a Landmark Trial

www.nytimes.com · 6 months ago

Snap has reached a settlement in a lawsuit alleging social media addiction, avoiding a trial that could have challenged the inherent design of social media platforms. The lawsuit, similar to cases against Big Tobacco, claims features like infinite scroll contribute to mental health issues in young users, and is one of several pending cases against major tech companies including Meta, TikTok, and YouTube. Plaintiffs are seeking monetary damages and design changes to curb excessive use. Snap's chief executive, Evan Spiegel, was expected to testify in the trial.

'CrashFix' Scam Crashes Browsers, Delivers Malware

www.darkreading.com · 6 months ago

A threat actor known as KongTuke is using a new technique called CrashFix to distribute malware. The scheme involves a malicious browser extension, NexShield, which intentionally crashes a user's browser before presenting a fake repair prompt. Corporate systems are targeted with ModeloRAT, a previously unseen remote access Trojan that collects system information and establishes persistence, while home users appear to be receiving a testing payload. Huntress Labs has published indicators of compromise to help organizations detect and prevent this evolving campaign.

Mass Spam Attacks Leverage Zendesk Instances

www.darkreading.com · 6 months ago

Multiple users are reporting receiving large volumes of spam emails originating from Zendesk domains, bypassing standard email filters. The spam messages often impersonate legal notifications or bogus lawsuits and aim to steal credentials or gain access to user accounts. Zendesk is investigating the issue, which may involve relay spam or vulnerabilities in system configurations, and has advised customers to implement specific security measures. ElevenLabs also experienced a mass spam attack impacting their Zendesk ticketing system, and Reliaquest reported potential preparations by threat actors linked to Scattered Lapsus$ Hunters targeting Zendesk environments.

EU Plans Phase Out of High Risk Telecom Suppliers, in Proposals Seen as Targeting China

www.securityweek.com · 6 months ago

The European Union is planning to phase out telecommunications equipment from companies in "high risk" countries over a three year period. This action is intended to address concerns over reliance on Chinese technology manufacturers like Huawei and ZTE. The proposed legislation would make cybersecurity measures mandatory across the EU, moving away from previous voluntary guidelines. Restrictions will extend beyond telecommunications to include sectors like border security, water systems, and medical devices.

AI-supported vulnerability triage with the GitHub Security Lab Taskflow Agent

github.blog · 6 months ago

GitHub’s Security Lab is using artificial intelligence to automate the process of analyzing security alerts, minimizing false positives and speeding up the identification of genuine vulnerabilities. The system uses large language models to gather information, audit alerts, generate reports, and create GitHub Issues for further investigation. Taskflows are broken down into modular steps for improved reliability and reusability, and this automation has already helped uncover approximately 30 real-world vulnerabilities, including cross-site scripting issues. The open-source framework allows others to build similar automated solutions.

VoidLink cloud malware shows clear signs of being AI-generated

www.bleepingcomputer.com · 6 months ago

Researchers have discovered VoidLink, a sophisticated Linux malware framework, likely developed with the assistance of artificial intelligence. Analysis of the developer's operational security failures revealed source code, documentation, and development plans, indicating the malware was produced primarily through AI-driven development within a week. The developer utilized an AI assistant and Spec-Driven Development to generate a multi-team development plan and code, resulting in a functional framework with approximately 88,000 lines of code by early December 2025. This marks the first documented case of advanced malware creation using AI, potentially enabling individual developers to achieve results previously requiring large teams.

Supreme Court to consider whether geofence warrants are constitutional

therecord.media · 6 months ago

false

New iOS and iPadOS Flaws Leave Millions of iPhones at Risk

www.techrepublic.com · 6 months ago

Critical security flaws have been identified in iOS and iPadOS, potentially affecting millions of devices. These vulnerabilities, CVE-2025-43529 and CVE-2025-14174, allow attackers to gain full access to a device without any user interaction. Apple has released a fix in iOS 26, but many users remain vulnerable due to older devices or failure to update, and targeted attacks are occurring. Users are strongly encouraged to update their devices to the latest iOS version to protect their data.

EU plans cybersecurity overhaul to block foreign high-risk suppliers

www.bleepingcomputer.com · 6 months ago

The European Commission is proposing new cybersecurity legislation to strengthen defenses against cyberattacks and state-sponsored groups. The legislation aims to remove high-risk suppliers from telecommunications networks, with concerns raised previously regarding companies like Huawei and ZTE. The new rules will allow the Commission to conduct EU-wide risk assessments and support restrictions on equipment used in critical infrastructure. This initiative seeks to enhance technological sovereignty and improve cybersecurity for all European citizens.

North Korea-Linked Hackers Target Developers via Malicious VS Code Projects

thehackernews.com · 6 months ago

North Korean threat actors are targeting developers using malicious Visual Studio Code projects as part of an ongoing campaign. Attackers instruct potential victims to clone repositories from platforms like GitHub and GitLab, then execute malicious payloads disguised within VS Code task configuration files. These payloads, hosted on Vercel domains, deploy backdoors like BeaverTail and InvisibleFerret, enabling remote code execution and data theft. The attackers are also employing fallback mechanisms, including malicious npm dependencies and JavaScript code, to increase the likelihood of successful compromise.

Billion-Dollar Bait & Switch: Exploiting a Race Condition in Blockchain Infrastructure

mavlevin.com · 6 months ago

A race condition vulnerability was discovered in the Ethereum network's auction system, allowing attackers to potentially win bids without paying the required fee. The vulnerability, found in the Relay's "UpdateTopBid" function, involved a timing gap between checking a bid and retrieving the associated data from Redis. Attackers could exploit this by submitting a high bid initially, then quickly replacing it with a zero-bid, effectively tricking the system into associating a low-cost payload with a high bid. The Flashbots team has implemented a patch using atomic operations to resolve the issue.

Sprocket Security Appoints Eric Sheridan as Chief Technology Officer

gbhackers.com · 6 months ago

false

Kimwolf Botnet Lurking in Corporate, Govt. Networks

krebsonsecurity.com · 6 months ago

A new botnet called Kimwolf has infected over 2 million devices, primarily Android TV streaming boxes, to carry out distributed denial-of-service attacks and relay malicious internet traffic. The botnet exploits vulnerabilities in residential proxy services, particularly those from IPIDEA, to scan local networks and compromise additional devices. Kimwolf has been found within government, corporate, and academic networks worldwide, with Infoblox reporting nearly 25 percent of its customers showing signs of Kimwolf-related activity. Security experts warn that a single compromised device can provide attackers with a foothold to probe and potentially compromise an organization's entire network.

Google Gemini Flaw Let Attackers Access Private Calendar Data

www.techrepublic.com · 6 months ago

Security researchers discovered a vulnerability in Google’s Gemini AI assistant that allowed attackers to extract private calendar data and create deceptive events. The flaw involved embedding hidden instructions within a meeting invite, bypassing Google Calendar’s privacy controls. Gemini’s ability to automatically process calendar data became the weak point, enabling attackers to trigger malicious instructions by prompting Gemini with seemingly normal questions. Google has since acknowledged the issue and implemented mitigation measures.

One Identity Unveils Major Upgrade to Identity Manager, Strengthening Enterprise Identity Security

gbhackers.com · 6 months ago

false

Vulnerabilities Threaten to Break Chainlit AI Framework

www.darkreading.com · 6 months ago

Researchers have identified two high-severity vulnerabilities in the Chainlit open-source AI framework, potentially exposing users to data loss and cloud account takeovers. The first vulnerability, CVE-2025-22218, allows attackers to access arbitrary files on a server by manipulating custom elements. The second, CVE-2026-21219, is a server-side request forgery bug that can be exploited to access sensitive systems. Experts warn that AI systems often combine multiple frameworks, leading to increased vulnerability and risk due to rapid development and overpermissioned access to data.