AI Safety Guidelines: 10 Essential Rules to Protect Data, Finances, and Reputation When Working with LLMs
A detailed analysis of emerging AI-related security risks highlights how large language models can autonomously execute attack chains, fall victim to prompt injection, and cause cascading errors in complex workflows. The article examines real-world incidents such as the Anthropic vending machine pricing failure, the Meta Instagram account takeover via overly helpful AI support, and Copilot Studio data leaks through prompt injection. It emphasizes that while attack methods themselves are not revolutionary, AI agents can now scale them at machine speed with autonomous decision-making and recovery capabilities. The piece provides ten concrete safety rules covering financial controls, fact verification, data confidentiality, context pollution prevention, and access limitation. It also stresses that ultimate responsibility always remains with the human operator, not the AI system.
Translated from Russian
Read full articleLatest News
From HAProxy to VLESS+Reality: Overcoming DPI Blocks for MTProto Telegram Proxies
A detailed case study describes the challenges of running an MTProto proxy for Telegram on Russian servers facing ISP-level DPI. Initial attempts using HAProxy TCP relays and SOCKS5 tunnels failed because modern DPI systems detect MTProto and fake-TLS signatures regardless of transparent forwarding. The author eventually succeeded by layering mtg with an Xray VLESS+Reality tunnel that performs genuine TLS 1.3 handshakes to legitimate domains. Key configuration pitfalls included the xtls-rprx-vision flow breaking non-TLS payloads, missing mux causing handshake timeouts on short-lived connections, and provider-specific network policies. Final architecture places mtg and an Xray client on the Russian entry server while the foreign exit server runs Xray in VLESS+Reality inbound mode. The guide supplies complete docker-compose examples, key generation commands, and a checklist of common misconfigurations to avoid.
Translated from Russian
TLS MITM and Throttling to 10 Mbps: Two Distinct Network Degradation Patterns Observed in Russia
Russian users have reported sudden slowdowns and instability on international connections over the past several days. Analysis of Tunnel Cat logs revealed two separate issues that produce similar user symptoms but require different diagnostic approaches. The first pattern involves successful TCP and TLS handshakes followed by sharp drops in throughput, often from 80 Mbps to 8-12 Mbps, consistent with DPI-based throttling of international traffic. The second pattern shows TLS certificate substitution on Windows systems, indicating active man-in-the-middle interception. Tunnel Cat now detects invalid certificates and terminates such sessions, yet the underlying network phenomena persist. The observations highlight the need to examine both bandwidth dynamics and certificate chains when troubleshooting connectivity problems from Russia.
Translated from Russian
CISA Adds Linux Kernel Frag Gap Flaw and Two Other Exploited Vulnerabilities to KEV Catalog
The US Cybersecurity and Infrastructure Security Agency (CISA) added three vulnerabilities to its Known Exploited Vulnerabilities (KEV) catalog on August 27, 2026. One of the entries is CVE-2026-53362, a high-severity privilege escalation issue in the Linux kernel also known as Frag Gap. The flaw stems from an out-of-bounds write when generating IPv6 packets, allowing a low-privileged user to obtain root access. The Linux Kernel Organization assigned it a CVSS v3.1 base score of 7.8 and rated it High severity, with public exploit code already available. Federal agencies must remediate CVE-2026-53362 and CVE-2023-49105 by the August 30 deadline. The advisory underscores ongoing exploitation of these issues in the wild.
Translated from Japanese
Aurora Ransomware Affiliate Uses AI Assistant Cursor to Compromise Active Directory and VMware ESXi Servers
An affiliate of the Aurora ransomware group employed the AI-powered coding assistant Cursor to plan and execute targeted attacks against corporate environments, focusing on Active Directory and VMware ESXi servers. The campaign impacted more than 20 organizations between April and July 2026. The operator used Cursor to generate commands, refine exploitation techniques, and build attack sequences aimed at Windows infrastructure, particularly Active Directory Certificate Services (ADCS). Techniques included noPac exploitation combined with NTLM relay attacks leveraging PetitPotam, PrinterBug, and DFSCoerce to escalate privileges up to domain administrator level. After gaining access, the attackers collected and compressed large volumes of data for exfiltration before deploying the Aurora ransomware on Windows, Linux, and ESXi systems. The ESXi variant stops virtual machines prior to encryption to maximize impact on virtualized environments.
Translated from Portuguese
Deleted Database Records Remain Recoverable in SQLite Files Despite DELETE Operations
A standard DELETE query in SQLite removes rows from the table view but leaves the actual data intact inside the database file until pages are reused. The pragma secure_delete setting controls whether freed pages are zeroed immediately or simply marked as available. With the default setting of 0, strings containing names and credit card numbers can still be extracted using grep even after deletion. The same behavior appears in PostgreSQL through dead row versions until VACUUM runs and in MySQL InnoDB through undo logs and the binary log. The issue directly affects compliance with personal data deletion requests because backups, replicas, and analytics exports often retain the original records. Proper mitigation requires enabling secure_delete, running VACUUM after bulk deletions, or encrypting sensitive fields with per-record keys that can be destroyed on request.
Translated from Russian
redb.Identity Adds gRPC Transport for OpenID Server Alongside Existing HTTP Facade
redb.Identity has introduced a second transport layer using gRPC next to its existing HTTP interface, sharing the same core routes, client registry, token store, and authorization logic. The new facade exposes standard OAuth and OpenID Connect operations such as Token, Introspect, Revoke, UserInfo, Discovery, and Jwks through protobuf-defined methods under identity.v1.Identity. Both transports enforce identical verdicts based on a single centralized scope table located behind direct-vm addresses, ensuring that a client authorized via HTTP receives the same result when calling gRPC. Error handling on gRPC uses status codes and trailers to carry machine-readable OAuth error codes and retry-after values, preserving compatibility with existing interceptors and tracing. Browser-facing flows, DPoP proofs, and user self-service remain on HTTP, while administrative operations are available on a separate management port. The implementation was validated through 64 unit tests, cross-language interop with @grpc/grpc-js clients, and a conformance run against the official OpenID Foundation suite.
Translated from Russian
Russian TSPU Systems Redirect DNS Queries to Google and Cloudflare Servers Toward National Domain Name System
Since the evening of August 26, Russian technical means of countering threats (TSPU) began intercepting open DNS queries sent to Google and Cloudflare public resolvers. For domains such as YouTube and RuTracker, UDP-based queries received NXDOMAIN responses while TCP queries successfully reached the original servers and returned valid IP addresses. Analysis with low TTL packets revealed that responses originated from IP address 195.208.5.1 belonging to the National System of Domain Names (NSDI). The mechanism performs targeted DNAT on recognized DNS traffic, making the query appear directed to NSDI rather than the foreign resolver. The redirection is imperfect, allowing subsequent identical queries sent in quick succession to bypass the system and reach Google. The findings come from experiments conducted by Habr user angry_agent and have not yet received official confirmation from Russian authorities.
Translated from Russian
Nvidia to Cease Regular GeForce Driver Updates for Windows 10 After October 2026
Nvidia has announced the end of regular driver support for Windows 10 in its GeForce Game Ready and Nvidia Studio driver lines starting October 2026. The first driver package without Windows 10 support will arrive in November of that year. Microsoft ended the base lifecycle of Windows 10 on October 14, 2025, and Nvidia is extending support by one additional year. Existing games and applications will continue to function after the change, while quarterly security patches for critical vulnerabilities will remain available until October 2029. Users will no longer receive optimizations for new games, fixes for graphics issues, or new GPU features. Newer technologies such as DLSS may also skip Windows 10 compatibility. The transition is described as gradual rather than abrupt, allowing older titles to keep running while newer releases increasingly encourage migration to a supported Windows version.
Translated from Russian
HTTPS Lock Icon Present but List of Visited Sites Remains Visible
Even when HTTPS is active and passwords stay protected, DNS queries and the SNI field in TLS handshakes expose the exact domains a user visits over public Wi-Fi. Classic unencrypted DNS over UDP sends domain names in plaintext, allowing anyone on the same network to observe them with simple packet captures. The SNI extension reveals the target hostname before encryption is negotiated, enabling domain-based filtering without decrypting traffic. DNS over HTTPS moves queries inside encrypted channels but shifts visibility to the chosen resolver instead of the local network. Encrypted Client Hello offers partial protection for SNI yet requires support from both browsers and server infrastructure. The practical takeaway is that metadata about services used, timing, and frequency leaks more readily than credentials in modern public networks.
Translated from Russian
From Russian sources
Translated from Russian
Russian TSPU Begins Intercepting UDP DNS Queries to Cloudflare and Google Public Resolvers
Starting on the evening of August 26, Russia's TSPU DPI system began actively intercepting plaintext DNS queries sent over UDP to public resolvers operated by Cloudflare and Google. Queries to 1.1.1.1 and 8.8.8.8 now return NXDOMAIN responses for blocked domains instead of the real IP addresses. The interception works exclusively on UDP; TCP-based DNS queries continue to receive legitimate answers from the original resolvers. Technical analysis shows the system performs targeted DNAT, rewriting the destination IP to the NSDI server at 195.208.5.1 only when a DNS query is detected inside the packet. Experiments with varying TTL values confirm that the redirection occurs after the traffic passes the TSPU node, and rapid successive queries can sometimes bypass the filter and return genuine records. The change affects netflow statistics visible to network operators, as traffic previously destined for foreign resolvers is now redirected domestically.
Rospotrebnadzor Introduces Age-Based Screen Time Limits for Russian School Students
Russia's consumer protection agency Rospotrebnadzor has established recommended maximum durations for schoolchildren working with computers and interactive whiteboards during lessons. The limits vary by grade, ranging from 20 minutes for first and second graders up to 35 minutes for students in grades 10 and 11. Separate rules apply to interactive boards, capping usage at 20 minutes for children under 10 and 30 minutes for older students. Schools must ensure students perform eye exercises when electronic devices are used, while traditional paper-based classes require such exercises only during breaks. Starting September 1 2026, a nationwide ban on mobile phones during lessons will also take effect, with individual schools deciding rules for recess periods.
Google Develops Public Android API for On-Device Content Safety Classification
Google is creating a new public Android API that will allow third-party applications to analyze images and other files locally on the device and assign them one of four safety statuses. The system builds on the existing SafetyCore component already used in Google Messages to blur intimate images. ContentSafetyManager will process images, raw files, and multimedia content entirely on-device without transmitting data to Google servers. Applications will then decide whether to display, blur, or hide the content based on the classification result. Experts warn that malicious apps granted broad permissions could misuse the classifier to scan large volumes of user files and build detailed profiles. SafetyCore itself previously sparked controversy after being installed automatically without explicit user consent and without a visible icon, leading some users to install blockers to prevent reinstallation via the Play Store. Google continues to emphasize minimal permission requests and local processing as safeguards for user privacy.
From Root CA to User Authorization in nginx and Apache: Client Certificate Login Explained
This is the third installment in a detailed tutorial series covering the deployment of a two-tier PKI infrastructure with Root CA and intermediate CAs for Person, Server, and Code. The article provides comprehensive configuration guidance for enabling mTLS in nginx and Apache, including full references for all ssl_client_* variables and SSL directives. It explains the differences between password-based and certificate-based authentication, the TLS handshake steps involving CertificateRequest and CertificateVerify, and the importance of proper extendedKeyUsage settings such as clientAuth. Readers learn how to issue client certificates, package them in PKCS#12 format, enforce revocation checks via CRL and OCSP, and safely pass certificate fields to backend applications while mitigating risks from header spoofing. The guide also covers scenarios where the application itself terminates TLS without a reverse proxy and demonstrates login flows protected against CSRF using one-time tokens.
Honeytoken Traps for Detecting Compromised Backends in Encryption Key Services
The article details a honeytoken-based detection system designed to identify when a trusted backend has been compromised and is abusing its signing key to request encryption keys for arbitrary documents. A secret set of fake user and document identifiers is stored only inside the key service using keyed HMAC-SHA256 hashes, ensuring the backend itself cannot discover or bypass the canaries. Any signed request touching these identifiers triggers an alarm, audit logging, and optional emergency lock that wipes master keys from memory. The check is deliberately placed after signature validation but before envelope decryption to avoid false positives from unauthenticated attackers while keeping performance cost low. Two types of canaries address different attack patterns: document canaries catch bulk scraping while user canaries detect forged grants for non-existent principals. The mechanism has been deployed in production, with live tests confirming silent failure responses externally and clear canary_trip plus emergency_lock events internally. Limitations include inability to catch targeted extraction of a single real userโs documents and the fact that detection occurs after a canary key has already been issued.
OSINT for the Lazy Part 19: AI as a Core Tool in Modern Intelligence Gathering
The article examines how artificial intelligence has transformed OSINT from a manual discipline into a scalable, automated process capable of handling massive data volumes. It details specific AI technologies including NLP models such as BERT, GPT and LLaMA for text analysis, computer vision tools like GeoSpy and Picarta for geolocation, and multimodal systems for processing mixed data types. Machine learning techniques for anomaly detection and Graph Neural Networks are presented as methods for uncovering coordinated campaigns and hidden networks. The piece also covers LLM agents that autonomously plan and execute multi-step OSINT tasks while stressing the continued necessity of human oversight for ethical judgment and verification. Limitations, ethical risks around privacy and attribution, and the growing asymmetry between state and independent actors are highlighted as critical concerns.
From Japanese sources
Translated from Japanese
Multiple Vulnerabilities Found in Apache Tomcat, Four Rated Critical by CISA
Apache Tomcat has received updates addressing 11 vulnerabilities across versions 11.0.25, 10.1.59, and 9.0.121. The Apache Software Foundation rated four issues as Important, while CISA assigned Critical severity to four CVEs based on CVSS v3.1 scores reaching 9.8. The flaws include authentication bypasses, access control evasion due to path evaluation order, off-by-one errors, and HTTP/2 resource leaks leading to denial of service. One vulnerability was disclosed earlier in July, with the remaining ten detailed on August 25. Moderate and Low severity issues were also patched in the same releases. The discrepancies in severity ratings between the vendor and CISA highlight differing risk assessments for the same CVEs.
Cisco Pre-Announces Security Advisories and Patches for Multiple Products on September 2, 2026
Cisco Systems has disclosed plans to publish security advisories for several product lines on September 2, 2026. The advisories will cover vulnerabilities affecting IP telephony devices, network switches, and email security appliances. Targeted products include Cisco IOS XR Software, multiple series of Cisco Desk Phones, Nexus 9000 Series switches with Silicon One, and Cisco Secure Email. The company will also provide updates aimed at strengthening security in IOS XR. No CVE identifiers, vulnerability details, affected versions, or CVSS scores have been released at the pre-notification stage. Cisco strongly recommends applying the forthcoming fixes once they become available, while noting that the schedule and product scope may still change.
CISA Adds Six Known Exploited Vulnerabilities Affecting NetScaler ADC, Linux Kernel and Microsoft SQL Server to KEV Catalog
The U.S. Cybersecurity and Infrastructure Security Agency has added six vulnerabilities to its Known Exploited Vulnerabilities catalog after confirming active exploitation in the wild. One of the flaws, CVE-2026-8452, affects Citrix NetScaler ADC and NetScaler Gateway products and can trigger denial-of-service conditions under specific configurations. The remaining five issues, disclosed between 2015 and 2022, impact the Linux Kernel, Red Hat Automatic Bug Reporting Tool, and the libuser library. Exploitation of these older flaws can allow local attackers to escalate privileges or corrupt password files. Organizations are urged to apply available patches and verify configurations immediately.
Google Releases Chrome 152 Fixing 327 Vulnerabilities Including 10 Critical Flaws
Google has released Chrome 152 for Windows, macOS, and Linux, addressing a total of 327 security vulnerabilities. Ten of these issues received the highest severity rating of Critical. The update resolves multiple Use After Free flaws in core components such as ANGLE, Aura, and Chromecast. Specific CVEs fixed include CVE-2026-79282 in ANGLE and several others in Aura and Chromecast modules. The release targets memory corruption and input validation weaknesses that could lead to remote code execution. Users are strongly advised to apply the update immediately to mitigate potential exploitation risks.
Veeam ONE Receives Security Update Addressing Critical Vulnerabilities Including CVE-2026-65641
Veeam Software has released updates for its backup environment monitoring tool Veeam ONE to address multiple vulnerabilities. The advisory covers the 13.x series and includes fixes rated as critical under CVSSv4.0. One flaw, CVE-2026-65641, allows service accounts to perform SMB authentication and carries a base score of 9.3. Additional high-severity issues such as CVE-2026-64633 with a perfect 10.0 score were also resolved in the patches. The company updated its disclosure to include seven vulnerabilities fixed in version 13.1.0.7034 after initially reporting six. Patches are now available for both the 13.x and 12.x branches, with the latest 12.3 Patch 1 backporting several fixes from the 13 series.
Three Critical Vulnerabilities in Adobe Campaign Classic Allow Arbitrary Code Execution
Adobe has disclosed three severe vulnerabilities in Adobe Campaign Classic that could permit attackers to execute arbitrary code on affected systems. The issues affect both Windows and Linux versions and include two OS command injection flaws along with one server-side request forgery vulnerability. Each vulnerability received the maximum CVSSv3.1 base score of 10.0 and is rated Critical. Adobe released fixes in version 7.4.4 build 9401, which must be applied to on-premises components in both pure on-premises and hybrid environments. Although no active exploitation has been observed, Adobe assigned the highest priority level and recommends applying the update within 72 hours.
From Portuguese sources
Translated from Portuguese
Next.js Patches Two Critical RCE Vulnerabilities in Versions 15.5.24 and 16.3.3
Next.js has released security updates to address two critical vulnerabilities that could allow unauthenticated remote code execution. The fixes are available in versions 15.5.24 and 16.3.3. The first issue, tracked as CVE-2026-75604 with a CVSS score of 9.0, affects applications hosted on Windows servers using Pages Router or App Router without Cache Components and stems from a path traversal flaw. The second vulnerability impacts the Image Optimization API when processing malicious AVIF files, enabling code execution through crafted image inputs. Both flaws affect a wide range of versions from 10.0.0 and 13.4 onward. Administrators are advised to update immediately, rebuild containers, and verify production environments run the patched releases, especially on Windows systems and those handling user-uploaded images.
Iran-Linked Tortoiseshell Group Deploys Malicious wtsapi32.dll Backdoor for Persistent Windows Access
Researchers have uncovered new tools deployed by the Tortoiseshell group, an Iranian-linked threat actor also tracked as Mirage Kitten, UNC1549, and Nimbus Manticore. The campaign features a Windows backdoor disguised as the legitimate wtsapi32.dll library that silently establishes reverse SSH tunnels over port 443 to maintain access to compromised networks. Active since at least 2018, the group has targeted defense, aerospace, technology, IT services, and military organizations primarily in the Middle East and the United States. The malware preserves expected Windows API functions while enabling command execution, file exfiltration, in-memory DLL loading, directory listing, and system reconnaissance. Associated infrastructure spans the United Arab Emirates, Saudi Arabia, the United Kingdom, Belgium, Canada, Australia, and Japan.
SonicWall Patches Critical Path Traversal and Update Flaws in NetExtender for Linux
SonicWall has released fixes for two high-severity vulnerabilities in its NetExtender client for Linux that could allow remote attackers to write arbitrary files with root privileges and manipulate the automatic update process. The flaws impact versions 10.3.5 and earlier, while the Windows version remains unaffected. CVE-2026-66152 carries a CVSS score of 8.8 and stems from improper handling of tar archives containing OPSWAT data, enabling path traversal that lets attackers escape the intended extraction directory. CVE-2026-66153 scores 7.0 and arises from inadequate symlink and temporary file handling in the NEService update mechanism. Both issues require user interaction to exploit, and no in-the-wild attacks have been observed so far. Administrators are urged to upgrade immediately to version 10.3.6 or later, as no workarounds exist.
Zscaler Fixes Multiple Critical Flaws in Client Connector Enabling RCE and Authentication Bypass
Zscaler has released patches for several vulnerabilities in its Client Connector agent that could lead to remote code execution, authentication bypass, local privilege escalation, and denial of service. The most severe issue, tracked as CVE-2026-59568 with a CVSS score of 9.1, allows unauthenticated remote attackers to execute arbitrary code within the context of the Zscaler Client Connector process. A second critical flaw, CVE-2026-59564 also rated CVSS 9.1, affects communication between the connector and its management portal, enabling attackers to circumvent authentication mechanisms. Additional vulnerabilities include a local buffer overflow on Android and ChromeOS tracked as CVE-2026-59566 with CVSS 8.4, as well as issues that could result in privilege escalation or service disruption. The flaws impact Client Connector versions across Windows, macOS, Linux, iOS, Android, and ChromeOS, with varying affected builds depending on the platform. Updated builds for lines 4.6 through 4.9 on Windows and equivalent fixes for other operating systems are now available.
WhatsApp Adds Support for Multiple Passkeys on Single Account
WhatsApp has expanded its account protection features by allowing users to register more than one passkey on the same profile. The update particularly benefits users who switch between Android and iOS devices while maintaining phishing-resistant login methods. Passkeys replace traditional codes and passwords with biometric authentication, facial recognition, or device lock mechanisms. Meta reports that more than one billion people already use the feature on WhatsApp, with Android support introduced in 2023 and iOS support added in 2024. The change reduces reliance on a single device during authentication by associating multiple keys with one account. Users can manage keys through Settings > Account > Passkeys. WhatsApp has also strengthened two-factor verification by allowing replacement of the six-digit PIN with longer passwords containing letters, numbers, and special characters.
AmnesiaStealer Malware for macOS Hijacks Authenticated Browser Sessions via ClickFix and Chromium
A new macOS malware strain named AmnesiaStealer has emerged that combines infostealing, persistence mechanisms, and the ability to silently take over already authenticated browser sessions. The infection chain begins with a fake GitHub page that tricks victims into copying a malicious command into Terminal using the ClickFix social engineering technique. Once executed, the Rust-based payload collects system data, displays a fake installer window to capture the user password, and attempts to unlock the Keychain for protected credentials. Targeted data includes browser cookies, history, extensions, documents, Apple Notes, Telegram sessions, and cryptocurrency wallet information. A secondary module copies the browser profile and launches a hidden Chromium instance controlled through the Chrome DevTools Protocol, allowing attackers to interact with active sessions that have already passed multi-factor authentication. Persistence is achieved via a disguised LaunchDaemon, and the malware attempts to remove installation artifacts afterward.
From Chinese sources
Translated from Chinese
Yisaqi Automation Magician Delivers Self-Healing RPA to Fix Low ROI in Enterprise AI Agents
Meituan's internal experiment revealed that widespread AI agent adoption drove daily token costs into the millions while introducing errors into business processes. MIT research confirmed that 95 percent of generative AI pilots fail to deliver measurable returns despite tens of billions in spending. Yisaqi's Automation Magician addresses these gaps by extending the agent lifecycle beyond conversation to include exploration, automated testing, self-healing, and export of stable scripts. The platform uses natural language input to generate resilient automation flows that adapt to UI changes without repeated model calls. Deployments at Jiage Food and a major central SOE financial shared service center demonstrate large-scale coverage of tax, finance, and manufacturing processes with minimal manual intervention. Zhongtian Technology reported 80 percent efficiency gains and annual savings exceeding one million yuan after rolling out the solution across 112 scenarios.
Redis Patch Bypass Enables Multiple RCE Exploits as PoCs for TLS and Stream Vulnerabilities Go Public
Multiple remote code execution vulnerabilities have been disclosed in Redis over the past month, including a critical patch bypass for CVE-2026-25243 that reintroduces Double Free flaws via crafted stream operations. The latest issue, QVD-2026-58458 affecting the TLS pending list, now has full technical details and a working PoC available, following the earlier QVD-2026-55651 disclosure. All three flaws impact nearly every production version still in use, from Redis 6.2.22 and below through 7.4.9 and 8.6.4. Although authentication is required, widespread weak or empty password configurations and exposed instances make exploitation trivial for attackers. Successful compromise grants arbitrary code execution in the Redis process context, enabling standard post-exploitation steps such as credential harvesting and lateral movement across internal networks. Organizations are urged to inventory all instances, apply the latest patches immediately, restrict network access via ACLs, and monitor for anomalous use of commands like XGROUP, EVAL, and RESTORE.
NVIDIA NemoClaw Flaw Lets Malicious Webpage Hijack Local Ollama Models via DNS Rebinding
Oasis Security disclosed a critical attack chain in NVIDIA NemoClaw that allows a malicious webpage to silently take over a local Ollama instance and poison AI model chat templates. The vulnerability stems from NemoClaw binding Ollama to 0.0.0.0:11434 on Windows without authentication, combined with skipped Host header checks and permissive CORS. Attackers use DNS rebinding to reach the local API from the browser and then inject persistent hidden instructions through the /api/create endpoint by modifying Go templates. These poisoned templates append attacker commands to every system message and survive across sessions and new prompts. No CVE has been assigned and no official patch exists, though version v0.0.106 added an incomplete bind check that can be disabled via environment variable. The issue revives a similar problem previously fixed in Ollama under CVE-2024-28224. Oasis Security notes this marks their third successful compromise of local AI agents using the same browser-to-local-API pattern.
Critical Zoom Vulnerability Exposes All Platforms to Remote Takeover via Screen Sharing Annotations
A high-severity vulnerability in Zoom Workplace clients across Windows, Mac, iOS, Android, and Linux allows attackers to remotely seize full device control simply by having a victim enable screen sharing and the annotation tool. The flaw requires no user interaction such as clicking links or dismissing warnings, and the compromise occurs silently without visible alerts. Security researchers demonstrated that AI tools enabled complete exploit development within 24 hours, dramatically lowering the barrier for advanced attacks previously limited to nation-state actors. The issue affects hundreds of millions of remote workers who rely on Zoomโs daily screen-sharing features. Official patches have been released, and users are urged to update immediately while avoiding annotation tools in sensitive sessions. Apple simultaneously addressed related macOS flaws in multiple versions. The incident highlights how generative AI is accelerating the weaponization of vulnerabilities.
NVIDIA Accelerates Physical AI Push with $500 Billion Infrastructure Plan at 2026 World Robot Conference
The 2026 World Robot Conference opened with 373 companies showcasing over 3,000 exhibits and more than 300 new products focused on embodied intelligence. NVIDIA is deepening its commitment to physical AI by partnering with Apollo, Blackstone, KKR and other major asset managers to create an independent financing platform targeting over $500 billion in third-party capital for AI infrastructure. Madison Huang, NVIDIAโs Senior Director of Physical AI and Robotics, visited the event to review advances in human data, simulation, and real-world deployment. The company is reframing compute resources as revenue-generating AI factories that produce tokens, simulation data, and action policies rather than treating them as cost centers. Humanoid robot development faces a critical bottleneck because high-fidelity physics simulation for millions of virtual agents demands enormous cloud compute, while sim-to-real transfer gaps and edge inference constraints remain unsolved engineering challenges.
Poisoned Rust Crates Execute Malware at Build Time: 245 Million Downloads Hit in Supply-Chain Attack
Three widely used Rust crates on crates.io were poisoned on August 20 with malicious versions that execute automatically during cargo build. The attack leveraged a typosquatted proc-macro1 dependency containing a build script that downloads payloads and establishes persistence. arrayref alone has accumulated 245 million downloads and is pulled automatically through caret ranges in many dependency trees. Attack infrastructure overlaps with prior campaigns attributed to Sapphire Sleet and MIDNIGHT NEPTUNE. Rust security teams yanked the malicious releases within 86-107 minutes, but the incident highlights missing publish-age controls and weak maintainer-account protections in the Cargo ecosystem.
From Spanish sources
Translated from Spanish
CISA Adds Six Actively Exploited Vulnerabilities to KEV Catalog Including Citrix NetScaler, Linux Kernel and Microsoft SQL Server Flaws
On August 26, 2026, CISA added six vulnerabilities to its Known Exploited Vulnerabilities catalog, signaling confirmed active exploitation and requiring immediate remediation priority. The batch includes a recent memory corruption issue in Citrix NetScaler ADC and NetScaler Gateway tracked as CVE-2026-8452, along with five older flaws affecting Microsoft SQL Server, the Linux kernel, Ajax.NET Professional, Red Hat libuser, and Red Hat ABRT. Citrix released patches for the NetScaler vulnerability on June 30, 2026, while CISA set an August 29, 2026 deadline for federal agencies. Real-world attacks have already deployed web shells and performed reconnaissance after successful exploitation of the Citrix appliance. The remaining CVEs enable remote code execution, local privilege escalation, and denial-of-service conditions across widely deployed enterprise technologies.
SLEEPWALKER Backdoor Activates on Windows via Single Custom Encrypted Packet
SLEEPWALKER is a passive Windows backdoor that stays dormant in memory until it receives one specially crafted encrypted network packet. Upon activation the implant decrypts and executes custom bytecode written in a proprietary 23-instruction language, eliminating traditional beaconing and reducing network indicators. The malware is delivered as a 64-bit DLL that impersonates dpapi.dll and is loaded via DLL side-loading into ERAAgent.exe, the executable of the ESET Management Agent used in ESET PROTECT deployments. It inspects raw traffic in promiscuous mode, supports multiple transport protocols including TCP, UDP, ICMP, SMB named pipes and VMware VMCI, and can weaken security by modifying registry values such as EveryoneIncludesAnonymous and NullSessionPipes. Configuration is protected with AES-256-CCM and the sample contains no confirmed victims or attributed infrastructure. Researchers have released YARA rules and read-only scanning utilities to detect the implant and its artifacts.
SLEEPWALKER Backdoor Stays Dormant Until Triggered by Single Custom Network Packet
SLEEPWALKER is a stealthy Windows backdoor that remains inactive until it receives one specially crafted encrypted network packet. The implant then decrypts and executes custom bytecode from a proprietary 23-instruction language, enabling chained tasks, data movement, and in-memory code execution without prior outbound communication. The 64-bit DLL masquerades as dpapi.dll and is designed for DLL side-loading into ERAAgent.exe, the executable of the ESET Management Agent used in ESET PROTECT deployments. It inspects raw network traffic in promiscuous mode, supports multiple transport protocols including TCP, UDP, ICMP, SMB named pipes, and VMware VMCI, and can weaken security by modifying registry values such as EveryoneIncludesAnonymous. No confirmed victims or attribution have been identified, and the analysis is based on a single sample. Researchers recommend auditing endpoints for unexpected dpapi.dll files alongside ERAAgent.exe and reviewing related registry settings.
Keycloak Fixes Critical CVE-2026-18963 Password Reset Flaw Allowing Unauthenticated Account Takeover
Keycloak has patched a critical vulnerability tracked as CVE-2026-18963 that lets remote unauthenticated attackers force a password reset for any user and seize control of the account. The flaw affects the reset-credentials flow due to improper state validation, enabling attackers to bypass email verification and reach the password change endpoint directly. With a CVSS 3.1 score of 9.1, the issue requires no privileges or user interaction and impacts confidentiality and integrity. Patches are available in Keycloak 26.7.2 and corresponding Red Hat Build of Keycloak releases 26.4.15 and 26.6.6. Organizations unable to update immediately should disable the Forgot password feature across all realms. Post-remediation auditing of login events and credential changes on privileged accounts is recommended to detect any prior exploitation attempts.
CISA Adds Actively Exploited Zimbra CVE-2026-73570 to KEV Catalog with Immediate Patching Deadline
CISA has added CVE-2026-73570 affecting Zimbra Collaboration Suite to its Known Exploited Vulnerabilities catalog, ordering urgent remediation due to confirmed active exploitation. The flaw enables unauthenticated remote command execution on servers where the optional zimbra-snmp package is installed and SNMP notifications are enabled. Exploitation occurs via specially crafted SMTP requests that trigger operating system command injection under the privileges of the zimbra user. Zimbra released the fix in version 10.1.20 on July 20, 2026, while CISA placed the vulnerability in the KEV list on August 21 and set an August 24 deadline for federal civilian agencies. Organizations are advised to first verify exposure by checking for zimbra-snmp and SNMP notification settings before applying the update. Post-patch steps include validating the version, controlled service restarts, and hunting for indicators such as unexpected Zimbra restarts, anomalous zimbra user activity, and new files in sensitive directories like /opt/zimbra/jetty/webapps/. The directive highlights the critical nature of pre-authentication flaws in internet-facing mail servers that process untrusted inbound traffic.
Critical GeoServer Vulnerability CVE-2024-36401 Actively Exploited for Unauthenticated Remote Code Execution
The CVE-2024-36401 flaw in GeoServer and its GeoTools library allows attackers to achieve remote code execution without authentication by abusing property name expressions interpreted as XPath. Active exploitation has already led to confirmed intrusions involving initial access, lateral movement, and persistence with tools such as China Chopper web shells. Multiple OGC endpoints including WFS GetFeature, WMS GetMap, and WPS Execute are affected when exposed to the internet. Patches are available in GeoServer versions 2.22.6, 2.23.6, 2.24.4, and 2.25.2. Organizations unable to patch immediately can mitigate risk by removing the gt-complex JAR file, though this may break functionality. Additional defenses include restricting internet exposure through IP allowlisting, VPNs, and reverse proxies while monitoring logs for anomalous requests. Any previously exposed instances should be treated as potentially compromised, with full incident response including credential rotation and host forensics recommended.