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
Dangling DNS Records Expose Companies to Subdomain Takeovers Years After Projects Shut Down
A marketing team requested reactivation of a three-year-old promotional landing page, only to discover that its domain now served an unrelated third-party website. The root cause was a forgotten CNAME record in the company's DNS zone that continued pointing to a project name on external hosting after the original service had been decommissioned. Because many hosting providers use wildcard DNS entries, standard availability checks and simple DNS resolution tests fail to detect that the target project no longer exists. Attackers who re-register the freed project name can then serve content under the company's legitimate subdomain, automatically obtaining valid TLS certificates and inheriting domain-wide cookies. The article details how such dangling records accumulate across infrastructure, marketing, and acquired company zones, and explains the concrete risks including cookie theft, bypass of Content-Security-Policy, and unauthorized domain ownership proofs. It concludes with a practical remediation workflow: export the full zone, assign owners to every record, validate external targets by application response rather than DNS resolution, and embed record removal into decommissioning procedures.
Translated from Russian
OTUS Publishes September Digest of Free Lessons on Linux Administration, PostgreSQL, CI/CD and Infrastructure Security
OTUS has released a new digest listing free September webinars aimed at infrastructure engineers, DevOps specialists and system administrators. The program covers practical topics including Linux server configuration, PostgreSQL 18 performance tuning, high-availability clusters with Patroni, CI/CD pipelines in GitLab, eBPF observability and infrastructure security practices. All sessions are delivered by practicing OTUS instructors who share real-world production experience. Separate tracks address RAID and LVM management, GPO policies, release management in 1C environments, Go profiling, mitmproxy traffic analysis and responsible use of AI tools for incident investigation and code review. The webinars run throughout September at 19:00 or 20:00 Moscow time and require only free registration. The digest also includes sessions on career growth from tech lead to CTO and effective responsibility distribution for team leads.
Translated from Russian
Google Patches Actively Exploited V8 Zero-Day in Chrome (CVE-2026-85046)
Google has released an urgent update for Chrome to address CVE-2026-85046, a high-severity zero-day vulnerability in the V8 JavaScript engine that is being actively exploited in the wild. The flaw is a type confusion issue within the Maglev compiler, specifically tied to optimizations around Array.prototype.sort, which can lead to arbitrary memory read and write primitives inside the JavaScript heap. Attackers can exploit the vulnerability by tricking victims into visiting a specially crafted HTML page, enabling remote code execution within the browser sandbox. The update brings Chrome to version 152.0.7977.82 or higher on Windows, macOS, and Linux, while also addressing a total of 12 vulnerabilities. The issue was reported by researcher Salvatore Gulizia, known as Serotav, who received a $1,000 reward. Other Chromium-based browsers such as Edge, Brave, Opera, and Vivaldi will require corresponding updates from their vendors.
Translated from Spanish
Google Patches Actively Exploited V8 Zero-Day in Chrome (CVE-2026-85046)
Google has released an urgent update for Chrome to address CVE-2026-85046, a high-severity zero-day vulnerability in the V8 JavaScript and WebAssembly engine that is being actively exploited in the wild. The flaw is a type confusion issue tied to the Maglev compiler and optimizations around Array.prototype.sort, which can lead to arbitrary memory read and write primitives inside the JavaScript heap. Successful exploitation allows a remote attacker to execute arbitrary code within the browser sandbox after a victim visits a specially crafted HTML page, a technique commonly seen in malvertising and compromised-site campaigns. The fix is included in Chrome 152.0.7977.82 and later for Windows, macOS, and Linux, and the update also resolves a total of twelve vulnerabilities. The vulnerability was reported on 4 August 2026 by researcher Salvatore Gulizia, known as Serotav, who received a $1,000 reward. Other Chromium-based browsers such as Microsoft Edge, Brave, Opera, and Vivaldi will require corresponding updates from their respective vendors.
Translated from Spanish
Top LLMs Misidentify Poisonous Mushrooms in Every Ninth Case, Benchmark Shows
Polish developer Piotr Migdal evaluated leading large language models on their ability to identify mushrooms from photographs, using a dataset of 1040 images covering 55 species common in Poland. The images came from the FungiTastic dataset derived from the Atlas of Danish Fungi, with expert labels and partial DNA confirmation. Models were asked to return the five most likely species names in Latin without additional training or tools. Gemini 3.8 Flash performed best with 65 percent top-1 accuracy and 85 percent top-5 accuracy, followed closely by other Gemini variants. However, safety-critical errors remained high: Gemini models labeled poisonous mushrooms as edible in roughly 11 percent of cases, while GPT-5.6 Sol reached 24 percent, Claude Opus 5 reached 29 percent, and Qwen 3.8 27B reached 36 percent. The study did not ask models directly whether a mushroom was edible; species identifications were later cross-checked against toxicity tables.
Translated from Russian
Over 15,000 Companies Deploy AI Agents to Combat Cloud Cyberattacks
More than 15,000 companies are already using AI-powered information security services in public clouds to detect attacks in real time, triage alert streams, and address vulnerabilities faster than human teams can process thousands of notifications. Yandex Cloud's threat report for the first half of 2026 shows attackers moving away from stolen passwords toward exploiting vulnerabilities in public applications. Specific attempts observed in Russian clouds include the critical React2Shell flaw along with Linux kernel issues Copy Fail and Dirty Frag. Companies are handing routine tasks such as alert prioritization, incident data collection, and initial investigations to AI agents while keeping human oversight for critical vulnerability remediation. The report notes that retail has become the top attack target at 39 percent, followed by manufacturing at 29 percent, while the IT sector dropped to 20 percent. Modern defenses now require behavioral detectors, unified telemetry, and AI agents to match the accelerated pace of attacks.
Translated from Russian
September 2026 AI Model Rankings: Fable 5.1 Tops Intelligence Index as Competition Tightens Across GPT-5.6 Sol, Grok 4.6 and Muse Spark 1.3
The beginning of September 2026 marked a rare moment when the list of top language models had to be almost entirely rewritten. Anthropic released Fable 5.1 and the limited Mythos 5.1, while Meta updated Muse Spark to version 1.3, Google introduced Gemini 3.8 Flash, and Alibaba refreshed Qwen3.8-Max. Existing models including GPT-5.6 Sol, Grok 4.6, Kimi K3, GLM-5.3 and DeepSeek V4 Pro remain competitive. Traditional rankings from smartest to least capable have become difficult because modern models operate in multiple reasoning-depth modes where low, high and max settings can differ by ten or more points on the same test. The market is better viewed as several overlapping races where Fable 5.1 leads in complex reasoning quality, GPT-5.6 Sol and Grok 4.6 deliver near-top performance at lower cost, and Muse Spark 1.3 excels in price-performance. Independent Artificial Analysis Intelligence Index scores, context windows, API pricing and tool-use capabilities now determine practical choices more than raw benchmark numbers.
Translated from Russian
How Russian Companies Can Legally Transfer Personal Data to Contractors Under 152-FZ
The article explains the legal distinction between data processors and independent operators when outsourcing tasks involving personal data. It details that the role of a contractor is determined by who sets the processing purpose, not by the service contract itself. For processors, a detailed data processing instruction under Article 6 of 152-FZ is required, while independent operators need a separate legal basis such as consent or contract performance. Special rules apply to employee data under Article 88 of the Labor Code, mandating written employee consent for transfers to third parties. The guidance also covers sub-processing risks, transparency obligations, and penalties under Article 13.11 of the Code of Administrative Offenses. Practical checklists help organizations classify contractors and prepare the correct documentation.
Translated from Russian
Drama RAT: Advanced Android Banking Trojan with RAT Capabilities Analyzed by Positive Technologies
Positive Technologies researchers have detailed Drama RAT, a sophisticated Android banking trojan distributed via phishing messages in messengers and disguised as VPN services or banking apps. The malware uses droppers that request installation of fake updates, then deliver a full-featured payload capable of stealing credentials, recording screens, and performing overlay attacks. Drama RAT automatically grants itself Accessibility permissions while hiding the process behind opaque overlays and employs multiple persistence mechanisms including AlarmManager alarms and Doze mode bypass. Communication occurs primarily over mTLS WebSocket with a fallback CDN channel, while anti-analysis features include over 1200 junk artifacts, broken AndroidManifest.xml, Frida hooking detection, and per-class string obfuscation. The trojan supports keylogging, SMS interception, VNC screen sharing, microphone and camera recording without indicators, and participation in DDoS attacks. Positive Technologies notes that removal is extremely difficult without ADB or safe mode because the malware redirects users away from settings screens.
Translated from Russian
From Russian sources
Translated from Russian
OpenAI GPT-6 Astra Reaches 100% Exploit Generation Accuracy on ExploitBench
OpenAI has unveiled GPT-6 Astra, its latest model that demonstrates unprecedented capability in generating working exploits from known vulnerabilities. The model scored a perfect 100% on ExploitBench, significantly outperforming the previous GPT-5.6 Sol at 78.5%. In tests involving recently disclosed vulnerabilities, Astra achieved arbitrary code execution in 39% of cases compared to 11.5% for its predecessor and discovered two previously unknown zero-day flaws. Due to these capabilities, OpenAI has restricted the public release to safe code analysis and patch development only, blocking requests for proof-of-concept exploits. Expanded access for vulnerability research and malware analysis will be provided through the Daybreak program to selected defenders, while the company adds stronger jailbreak protections and activity monitoring.
InfoWatch Acquires Web Control DC Team and Rebrands sPACE PAM as InfoWatch Privilege Control
InfoWatch has expanded its product portfolio by incorporating the Web Control DC development team and rebranding its flagship sPACE PAM solution. The new product, InfoWatch Privilege Control, is designed to manage and monitor privileged accounts belonging to system administrators, contractors, external specialists, and business users. These accounts provide access to servers, databases, network equipment, and critical applications, making them high-value targets for attackers. According to InfoWatch data, approximately 40% of critical information security incidents in Russia in 2025 were linked to the leakage or misuse of privileged credentials, while another 30% of confirmed cyberattacks occurred through compromised IT contractors. The solution enables time-limited privilege issuance, connection management, and detailed activity logging to prevent unauthorized actions. The original sPACE PAM product remains listed in the Russian software registry and holds FSTEC Russia certification at the fourth trust level, with compatibility for Astra Linux, Alt, and RED OS operating systems.
Secure Custom Domain Setup for Client Status Pages Using CNAME, Certbot and Go Instead of ACME On-Demand
A detailed case study describes how a monitoring service implemented white-label status pages on customer domains without relying on ACME on-demand certificate issuance. The approach uses pre-validated domains stored in a database, background DNS checks via CNAME or A-record matching, and a root helper script running on a systemd timer to handle certbot issuance and nginx configuration. Key design choices separate privileges so the Go application never touches certificates or nginx directly, while loopback endpoints are protected against proxy header spoofing. The solution explicitly addresses risks such as DoS through malicious Host headers exhausting Let's Encrypt rate limits, private key exposure, and first-visitor latency during TLS handshakes. Hysteresis in DNS status prevents temporary resolution glitches from disabling active customer pages. The entire implementation stays under a few hundred lines of code and runs reliably on a single server with nginx in front of the Go application.
Durev VPN Accused of Plagiarizing Independent Researchers' Articles for Commercial YouTube Promotion
Independent Russian cybersecurity researcher zarazaexe has publicly accused the commercial VPN service Durev VPN of systematically copying technical articles about Russian internet censorship systems, messenger analysis, and government-issued certificates. The team behind Durev VPN allegedly rewrote the original research into YouTube video scripts, replaced first-person statements with references to their own specialists, and removed all attribution to the source author or project. One video titled СРОЧНО УДАЛИ СЕРТИФИКАТ МИНЦИФРЫ reportedly gained 795,000 views in two days, while the service's Telegram bot shows 260,000 users and charges a minimum subscription of 313 rubles per month. Specific copied elements include scanning results of 46 million Russian IP addresses that yielded 63,000 entries in TSPU whitelists, architectural descriptions of default-deny and fail-closed policies, ECH handling by DPI systems, and detailed reverse-engineering findings from the MAX messenger client. The researcher documented the reuse of his earlier errors about UDP blocking inside whitelists and identical analogies comparing root certificates to apartment keys. When confronted, the Durev VPN representative initially demanded patents, invoked fair use, and later claimed the matter would be reviewed by their lawyer within a week. The researcher published the full chat logs and removed the disputed segment from one video after the complaint.
Protestware: How Ideologically Motivated Attackers Compromise Open Source Code and Supply Chains
Protestware represents a growing supply chain risk where open source maintainers deliberately insert destructive or politically motivated code into widely used packages. Unlike traditional malware seeking financial gain, protestware is driven by ideological motives and can delete files, cause denial of service, or restrict functionality based on geography or locale. Notable incidents include the node-ipc and peacenotwar packages that overwrote files on systems in Russia and Belarus, the colors.js infinite loop causing DoS, and es5-ext triggering high CPU usage during specific hours or Russian locales. These attacks often hide in postinstall hooks, use geolocation checks via ipinfo.io, or activate via environment variables and timestamps. Detection requires static analysis for geo-IP logic, dynamic monitoring in sandboxes, and tools such as Software Bill of Materials and OpenSSF Scorecard. Organizations are advised to adopt Zero Trust for dependencies, pin versions, use internal artifact repositories, and follow frameworks including NIST SSDF and SLSA to mitigate risks.
Bcrypt Password Hashing Silently Ignores Characters Beyond 72 Bytes, Breaking Verification Expectations
The bcrypt algorithm, widely used for password hashing in PHP and Python applications, processes only the first 72 bytes of any input password. Additional characters are ignored without error or warning during both hashing with password_hash and verification with password_verify. This behavior stems from the fixed 72-byte P-table size in the underlying Blowfish cipher as defined in the 1999 specification. Long passphrases, especially those using multibyte UTF-8 characters like Cyrillic or emojis, can result in completely different strings producing identical hashes. The limitation creates risks during password manager usage, system migrations, and scenarios involving shared prefixes. Modern libraries such as Python's bcrypt 4.x now explicitly reject passwords exceeding 72 bytes, while older implementations continue silent truncation. Recommended fixes include enforcing byte-length validation or pre-hashing with SHA-256 before bcrypt.
From Japanese sources
Translated from Japanese
Critical Vulnerabilities Patched in VMware Workstation and Fusion Allowing Host Code Execution
Broadcom has disclosed two vulnerabilities in VMware Workstation and VMware Fusion rated as critical and important respectively. CVE-2026-59346 is an integer overflow flaw in the VMXNET3 virtual network adapter that can be exploited by a local administrator inside a virtual machine to execute arbitrary code on the host system, carrying a CVSS v3.1 base score of 9.3. CVE-2026-59347 is a stack-based buffer overflow in the HGFS file system that permits code execution on the host as the VMX process with a CVSS score of 8.1. Both issues were privately reported and affect users who run virtual machines with local administrative privileges. The flaws have been addressed in VMware Workstation 26H1u1 and VMware Fusion 26H1u1, and Broadcom urges immediate updates via advisory VMSA-2026-0007.
Google Chrome Patches 12 Vulnerabilities Including Actively Exploited V8 Zero-Day
Google has released security updates for Chrome addressing 12 vulnerabilities across Windows, macOS, and Linux platforms. The updates include versions 152.0.7977.83 and 152.0.7977.82 for Windows and macOS, and 152.0.7977.82 for Linux. No critical-rated flaws were fixed in this batch, but ten issues carry a high severity rating. The most notable is CVE-2026-85046, a type confusion flaw in the V8 JavaScript engine that was reported on August 4 and is already being exploited in the wild. Additional high-severity fixes cover a race condition in V8 (CVE-2026-85045), a use-after-free in Compositing (CVE-2026-85048), and another type confusion issue (CVE-2026-85051). This marks the second Chrome security update in three days.
CISA Adds Seven Exploited Vulnerabilities in SonicWall, JFrog and Kestra to KEV Catalog
The US Cybersecurity and Infrastructure Security Agency has added seven known exploited vulnerabilities affecting SonicWall SMA1000, JFrog Artifactory and Kestra OSS products to its KEV catalog. Five of the flaws carry a three-day remediation deadline for federal agencies. Two issues in SonicWall SMA1000 enable unauthenticated server-side request forgery and authenticated OS command execution. A critical authentication bypass in JFrog Artifactory allows remote attackers to obtain administrative privileges. Kestra OSS is affected by an OS command injection vulnerability that can lead to full system compromise. CISA urges immediate patching and incident response actions.
Cisco Issues Security Updates for IOS XR Fixing Multiple Critical Vulnerabilities
Cisco Systems has released security updates for its network operating system Cisco IOS XR to address multiple vulnerabilities discovered through internal research. The update, published on September 2, 2026, classifies the flaws into seven categories, each assigned distinct CVE identifiers. Two vulnerabilities, CVE-2026-20274 and CVE-2026-20279, received the highest CVSS v3.1 base score of 9.8 and are rated Critical. Additional issues include command injection flaws rated 8.8 and other weaknesses scoring between 8.2 and 8.6. No exploitation has been observed at the time of disclosure. Cisco is distributing Software Maintenance Updates and plans to include fixes in upcoming releases 26.3.1 and 26.2.2.
HPE Aruba Networking Fabric Composer Affected by 52 Vulnerabilities Including Five Critical Flaws
HPE disclosed a security advisory on September 1, 2026, revealing 52 vulnerabilities in HPE Aruba Networking Fabric Composer. Five of these issues are rated Critical, including authentication bypass flaws that allow remote code execution and privilege escalation. CVE-2026-76657 permits attackers to bypass API authentication and gain administrator rights remotely. CVE-2026-76658 enables unauthenticated command execution as a privileged user via the SSH daemon. Additional critical issues include CVE-2026-19766 for adjacent network host takeover, CVE-2026-73700 for XSS in the web interface, and CVE-2026-73701 for unauthenticated remote code execution under certain conditions. HPE strongly recommends immediate updates across affected components including the OS, SSH, web management, and API layers.
SonicWall SMA 1000 Series Affected by Critical Vulnerabilities with Evidence of Exploitation
SonicWall has disclosed two serious vulnerabilities in its SMA 1000 series remote access appliances. The issues, tracked as CVE-2026-83548 and CVE-2026-83549, affect the SMA 6210, SMA 7210, and SMA 8200v models. CVE-2026-83548 is a server-side request forgery flaw in the WorkPlace interface that permits unauthenticated remote access to otherwise restricted functions. CVE-2026-83549 is an OS command injection vulnerability in the Appliance Management Console that can lead to remote command execution after authentication under certain conditions. Both vulnerabilities received high CVSS scores, with one reaching the maximum of 10.0. The vendor published the advisory on September 1, 2026, and confirmed cases where the flaws appear to have been exploited in the wild.
From Portuguese sources
Translated from Portuguese
Pegasus Spyware Returns in Serbian Surveillance Campaign via Zero-Click iMessage Exploit
A Serbian student activist's iPhone was infected with the Pegasus spyware through a zero-click exploit in iMessage, allowing silent installation without any user interaction. The infection, confirmed by Citizen Lab in collaboration with the SHARE Foundation, showed indicators of compromise between December 2025 and January 2026. Apple later sent the target a notification warning of a mercenary spyware attack attempt. The exploit granted full access to photos, messages, files, and enabled covert microphone and camera activation. The vulnerability was addressed in the iOS 18.4.1 update released on April 16, 2025. The incident forms part of a wider surveillance wave in Serbia, with at least 14 individuals including students, activists, a parliament member, and a local political representative receiving similar Apple alerts. Additional targets were hit with Android spyware variants linked to NoviSpy.
Attackers Abuse Legitimate Node.js Runtime to Deploy Persistent Malware in Targeted Operations
Cybercriminals are increasingly misusing the legitimate Node.js runtime to execute malicious JavaScript scripts and maintain persistence on corporate networks. The technique has been observed since February 2026 against governments, technology firms, hotels, and financial organizations. Attackers favor the signed node.exe binary to evade signature-based detections after initial access via ClickFix social engineering. In one case against an Asian technology company, Node.js was deployed following blocked attempts to install AdaptixC2 and Cobalt Strike. The resulting implant establishes registry persistence and communicates with Ethereum blockchain gateways to retrieve commands hidden in smart contracts. The approach has been linked to multiple tools including ModeloRAT, Mistic, EtherRAT, and a variant of AsukaStealer, with some campaigns attributed to the initial-access broker Woodgnat, also known as KongTuke.
Dropbox Lenovo ID Integration Flaw Exposes Thousands of Accounts Without Passwords
A flaw in the federated authentication process between Dropbox and Lenovo ID allowed attackers to compromise approximately 5,000 accounts between August 4 and August 21, 2026. Attackers registered a Lenovo ID using the email address tied to an existing Dropbox account due to missing email verification checks. They then logged in via the Lenovo ID option, and Dropbox automatically associated the identity with the existing account without requiring the original Dropbox password. Only accounts linked to the Lenovo ID integration and lacking two-factor authentication were affected. In fewer than one-third of the compromised accounts, attackers viewed or downloaded stored files. Dropbox responded by terminating all Lenovo ID sessions, removing existing associations, and updating the login flow to require the Dropbox password. Lenovo described the issue as stemming from a legacy integration and stated that its own customers were not directly impacted.
Ransomware Operators Linked to The Gentlemen Deploy TukTuk C2 Framework for Espionage and Credential Theft
Operators associated with the ransomware group The Gentlemen have adopted a new command-and-control framework called TukTuk to steal credentials, monitor compromised systems, and prepare environments for ransomware deployment. The framework was discovered on a server that also hosted tools for disabling EDR solutions, research on vulnerable drivers, and data apparently stolen from two large organizations. TukTuk includes agents for both Windows and Linux, along with its own backend and management panel that allows remote command execution, file transfers, screen capture, and device tracking through a single interface. One notable feature displays a fake Windows Security window on the victim's machine to capture entered credentials and send them directly to the attackers' panel. Researchers identified a DLL sideloading technique that abuses the legitimate Greenshot.exe executable to load a malicious log4net.dll library and launch the TukTuk agent. The server also contained EDRKiller, WarsawKiller, and UnknownKiller tools, plus materials on BYOVD attacks that leverage vulnerable legitimate drivers to gain kernel access and interfere with security products.
Password Spraying Campaign Targets AWS Root Accounts in Over 150 Organizations
A password spraying campaign targeted AWS root accounts across more than 150 organizations between July 24 and August 23, 2026. Attackers performed repeated login attempts against identities holding maximum privileges in the cloud environment. The root account is created with every AWS account and grants full access to resources, configurations, billing, and sensitive administrative functions. Researchers observed a median of two attempts per organization, with some targets receiving up to eight attempts. No successful authentications linked to the campaign have been identified so far. The attacks leveraged distributed proxies across multiple countries and networks, including hosting infrastructure and residential proxies, while using user agents that mimicked older versions of Microsoft Edge and Firefox. Since June 2025, AWS has required MFA for root users, significantly raising the bar for account takeover even if a password is discovered.
Brazilian Malware BraZetsu Uses AI to Map and Sell Compromised Corporate Networks
Researchers have identified a new Windows malware called BraZetsu developed by the Brazilian actor Exilware. The tool maps compromised corporate networks, identifies high-value systems, and prepares access for sale to other criminals. The operation primarily targets Brazil and other countries in Latin America and the Iberian Peninsula. BraZetsu is written in Python with a modular architecture and maintains WebSocket communication with its command-and-control server. It searches for banking software, ERPs, e-commerce platforms, industrial environments, development tools, backup systems, and security solutions. The malware also looks for Brazilian CNAB financial files, digital certificates, and browser history. Evidence suggests heavy use of generative AI during development, with an AI component on the server that may classify stolen files and prioritize valuable machines.
From Chinese sources
Translated from Chinese
CISA Adds Seven Actively Exploited Vulnerabilities to KEV Catalog, Including SonicWall and LiteLLM Flaws Targeting AI Infrastructure
CISA has added seven vulnerabilities confirmed to be exploited in the wild to its Known Exploited Vulnerabilities catalog, with two carrying CVSS scores of 10.0. The list includes critical issues in SonicWall SMA 1000 devices, JFrog Artifactory, Kestra workflow engine, and LiteLLM AI gateway components. Attackers are now systematically targeting AI infrastructure to steal large language model API keys and configuration data from databases such as LiteLLM_ProxyModelTable. Qilin ransomware operators have already chained LiteLLM flaws for remote code execution and data theft. CISA has set a 48-hour remediation deadline for federal agencies, underscoring the urgency for all organizations running these components.
Zhou Hongyi Warns AI Tools Are Industrializing Vulnerability Discovery
At the Fourth Cyberspace Security Forum in Tianjin, 360 founder Zhou Hongyi stated that vulnerability mining is shifting from artisanal workshops to automated production lines, compressing discovery cycles from months or years down to hours. AI tools such as Mythos are standardizing and automating the process, enabling attackers to replicate elite hacker expertise at scale through distilled models and agent swarms. 360's own Tulongfeng platform has already discovered over 10,000 vulnerabilities since its June release, including long-hidden high-risk flaws in Windows, Office, OpenClaw, Flowise, and Codex. The emergence of multi-agent systems introduces new attack surfaces because compromised agents can autonomously collaborate and move laterally faster than human operators. Zhou described this as the "second one-way transparency," where offensive tradecraft becomes copy-pasteable via prompts and toolchains. Defenders are advised to adopt "model-versus-model" strategies, automate vulnerability intelligence workflows with SOAR, enforce strict agent permission audits, and integrate AI into their own code review and detection engineering processes.
Anthropic Fable 5.1 System Prompt Fully Leaked Hours After Launch Exposing 275000 Characters of Rules
Anthropic released its flagship Fable 5.1 model alongside Mythos 5.1 on September 2, achieving strong benchmark scores including 90 percent on ARC-AGI-2. Within hours, researcher Pliny the Liberator published the complete 275000-character system prompt on GitHub, far exceeding the company's official 27000-word disclosure. The leaked document details 46 built-in tools, strict copyright restrictions, memory classification boundaries, and behavioral constraints that function as an internal employee handbook. The incident highlights that model weights remain the true core while prompt-based guardrails create an attack surface once mapped. It also reveals privacy rules that permanently exclude storage of minor identities, criminal records, and self-harm indicators even when users disclose them. The leak underscores the growing gap between vendor transparency claims and actual runtime instructions governing frontier AI systems.
Aiker World AI Co-Creation Base Officially Lands in Hainan Oriental to Build Long-Term AI Talent and Industry Ecosystem
On August 28, the finals of the Oriental Chuxin FC-OPC Next iCreate AI Native Content Creation Competition were held in Hainan Oriental. Aiker World community, in partnership with Oriental Chuxin, advanced their collaboration by establishing an offline AI co-creation base and connecting online talent with real industrial scenarios. Thirty youth teams presented AI projects focused on Chinese cultural dissemination, Hainan tourism and intangible heritage export, free trade port digital economy, and AI content compliance. The event was attended by senior local officials including Oriental Municipal Party Secretary Li Aihua and Hainan Provincial Tourism and Culture Department Deputy Director Wang Zhongyun. 360 Group representative and Aiker World initiator Hu Xiaona highlighted the shift from AI model development to practical enterprise deployment using tools like Nami Work. The new base aims to link talent, experts, and companies for sustained project delivery beyond single events.
1000 AI Builders Gather for 96-Hour SheNicest Hackathon in Beijing to Build Working Demos
Nearly 1000 creators from technology, product, design, hardware and imaging fields have assembled at the Beijing National Convention Center for the SheNicest 2026 Summer Thousand-Person Hackathon. Over 96 hours participants form teams, iterate rapidly and attempt to convert initial ideas into functional, demonstrable AI products across native applications, agents, intelligent hardware, games and AIGC imagery. A workshop led by Hu Xiaona of 360AI Business Development, HackingClub and AikerWorld examined the transition from individual AI tool use to full organizational adoption. She outlined three progressive stages: tool empowerment, workflow automation and true AI-native organizational redesign. The discussion stressed that simply purchasing AI tools does not change outcomes unless processes, roles and collaboration mechanisms are restructured around human-plus-AI teams. The event concludes with Demo Day on August 30, after which builders must address sales, delivery and sustained value creation.
ServiceNow Discloses Three CVSS 10.0 Vulnerabilities Allowing Unauthenticated Remote Code Execution and SQL Injection
ServiceNow has released security updates addressing four vulnerabilities in its AI platform, including three rated CVSS 10.0. The flaws enable unauthenticated attackers to achieve remote code execution, privilege escalation, and arbitrary SQL execution against core ITSM systems used by large enterprises worldwide. Affected components include the GraphQL Composite Data API, system configuration image upload processor, and dynamic schema ORDER BY handling. ServiceNow states it has patched hosted instances and provided hotfixes for self-hosted customers running Xanadu, Yokohama, Zurich, and Australia branches. This follows a July disclosure of a related sandbox escape tracked as CVE-2026-6875 that showed signs of in-the-wild exploitation. No public exploits or confirmed active attacks have been observed for the new issues yet, but the extremely low attack complexity leaves a narrow remediation window for organizations running exposed instances.
From Spanish sources
Translated from Spanish
Attackers Exploit Critical JFrog Artifactory Flaw CVE-2026-82329 to Forge Admin Tokens and Poison CI/CD Pipelines
A critical vulnerability tracked as CVE-2026-82329 in JFrog Artifactory is being actively exploited in the wild, allowing unauthenticated attackers with network access to generate or forge administrator tokens on self-managed installations that retain default settings. The flaw enables full authentication bypass, granting attackers administrative privileges to enumerate users and groups, access sensitive configurations, read stored artifacts, and modify security settings. The greatest risk arises when Artifactory serves as the central repository for automated CI/CD pipelines, where an attacker can replace trusted internal packages, container images, or dependencies and propagate malicious changes downstream into build runners and production environments. JFrog has released patches for multiple branches, including versions 7.111.21, 7.117.28, 7.125.20, 7.133.29, 7.146.38, and 7.161.20, although patching alone does not invalidate already-issued tokens. Organizations are advised to restrict network exposure, revoke and rotate admin tokens, review audit logs for anomalous token generation and configuration changes, and validate artifact integrity across the vulnerable period. The issue does not affect JFrog Cloud instances.
Critical JFrog Artifactory Flaw CVE-2026-82329 Actively Exploited to Forge Admin Tokens in Self-Managed Deployments
A critical vulnerability identified as CVE-2026-82329 in JFrog Artifactory is being actively exploited in the wild, enabling unauthenticated attackers to generate or forge administrator tokens simply by having network access to affected instances. The flaw impacts self-managed installations that retain default configurations, a common scenario when repositories are deployed quickly and later exposed to the internet. Once administrative access is obtained, attackers can enumerate users and groups, access sensitive configurations, read stored artifacts, and modify security settings. The most severe risk arises when Artifactory automatically feeds CI/CD pipelines, allowing malicious substitution of trusted packages, container images, or dependencies that then propagate downstream into production environments. JFrog Cloud instances were already protected, while self-managed deployments require updates to specific patched versions including 7.111.21, 7.117.28, 7.125.20, 7.133.29, 7.146.38, or 7.161.20. Even after patching, previously issued tokens must be revoked because their validity operates independently of the software update. Organizations are advised to restrict network exposure, rotate credentials, audit token generation logs, and validate artifact integrity for any material published during the vulnerable period.
Attackers Exploit Critical Sangoma Switchvox SQL Injection Flaw CVE-2026-9586 to Deploy Reverse Shells Without Credentials
A critical vulnerability in Sangoma Switchvox allows unauthenticated remote code execution through SQL injection in the /pa HTTP endpoint. The flaw, tracked as CVE-2026-9586 with a CVSS score of 9.3, affects internet-exposed VoIP deployments common in SMB environments. Attackers have already leveraged the issue in active campaigns since August 30, 2026, to open reverse shells and perform reconnaissance using Base64-encoded commands. The root cause is improper handling of the PhoneIP parameter in PostgreSQL queries, classified as CWE-89, which grants superuser database privileges. Sangoma released the fix in Switchvox 8.4.0.2 on July 14, 2026, following initial disclosure in April. Administrators are urged to patch immediately, restrict access, and monitor for indicators such as connections to 176.65.148.184 and suspicious entries in /var/log/switchvox/db-quirks.log.
Attackers Exploit Critical Langflow and Ruby on Rails Flaws for Credential Theft and C2 Infrastructure
Active exploitation has been confirmed for CVE-2026-0768 in Langflow and CVE-2026-66066 in Ruby on Rails. Attackers first focus on reconnaissance and secret exfiltration before establishing command-and-control channels, with potential escalation to remote code execution. In Langflow the flaw allows arbitrary Python code execution as root due to insufficient input validation, enabling attackers to dump environment variables and locate cloud credentials. The Ruby on Rails issue, tracked as KindaRails2Shell, stems from an arbitrary file read triggered when Active Storage processes untrusted image uploads with libvips, exposing secret_key_base and other sensitive keys. Observed campaigns show traffic originating from Russia against Langflow instances and activity targeting canaries in Singapore, Israel, and the United Kingdom for Rails deployments. Defenders are advised to inventory public instances, apply the latest patches, rotate exposed secrets, and monitor for suspicious file reads and C2 communications.
Over 8,300 Gitea Servers Remain Exposed to Remote Code Execution via CVE-2026-60004
Thousands of publicly accessible Gitea instances were still unpatched against CVE-2026-60004 as of late August 2026, leaving them open to remote command execution. The critical vulnerability combines abuse of the diffpatch endpoint with the ability to install and trigger attacker-controlled Git hooks, allowing unauthenticated attackers to run commands on the server. Exploitation requires only write access to a repository, which is often trivial due to default open user registration in many deployments. The issue was fixed in Gitea 1.27.1 released on July 27, 2026, and CISA added the flaw to its Known Exploited Vulnerabilities catalog with a remediation deadline of August 28 for U.S. federal agencies. In observed incidents, attackers used the access to deploy cryptocurrency miners inside Docker containers. A related authentication bypass affecting the official Docker image, CVE-2026-20896, was also noted when reverse-proxy headers such as X-WEBAUTH-USER are enabled.
CISA Adds CVE-2026-8452 to KEV Catalog and Orders Federal Agencies to Patch Citrix NetScaler by August 29 2026
CISA has placed CVE-2026-8452 in its Known Exploited Vulnerabilities catalog, compelling US federal agencies to apply patches to Citrix NetScaler before the unusually tight deadline of August 29 2026. The flaw, initially reported as a memory overflow leading to denial of service, has been confirmed through public technical analysis to enable unauthenticated remote code execution with root privileges on unpatched devices. The vulnerability impacts NetScaler ADC and NetScaler Gateway appliances configured as VPN Gateway or AAA virtual servers, with heightened risk in SAML-based deployments acting as either Service Provider or Identity Provider. Attackers have already conducted spray-and-pray campaigns that deploy webshells and perform basic reconnaissance after initial compromise. Tens of thousands of NetScaler instances remain internet-exposed, increasing the likelihood of automated mass exploitation. Citrix has released fixed builds including 14.1-72.61, 13.1-63.18, and 13.1-37.272 or higher for FIPS and NDcPP environments. Organizations are advised to inventory exposed devices, verify versions, hunt for indicators of compromise, and apply additional hardening measures beyond patching.