# New Wave of Cyber Threats: AI-Powered Attacks and Vulnerabilities From China-linked cyber operations to npm package risks, this week's critical security exposures. - **Source:** [The Hacker News](https://brifing.si/source/f944ec2b-8180-47bc-94de-57085b5a6563) - **Published:** 2026-08-06 - **Topic:** [Business](https://brifing.si/c/business) - **Original article:** - **On Brifing:** https://brifing.si/s/9e234079-f9d0-4a50-ab77-df95ecc4aac5 ## China's Hidden Telecom Infrastructure Threat The 49-page 'Stranger Pings' report by the U.S. Congress reveals risks posed by Chinese-controlled telecom companies in U.S. communication infrastructure. These companies allegedly serve as backdoors for state-sponsored cyber operations. Contractual 'Acceptable Use Policies' prohibit content violating China's state security laws. Campaigns like Salt Typhoon could be facilitated through this infrastructure. ## Rust-Based Backdoors in PDFs The SideWinder hacking group distributes new Rust-written backdoors through phishing PDFs. The malware establishes persistence via registry modifications and receives remote commands through free server platforms like Cloudflare Workers. Designed as part of multi-stage attack chains, it collects system information from victims before executing follow-on attacks. ## 846 Malicious Packages on npm The 'Flooding Dropper' campaign published 846 malicious packages on npm. Attackers automate account creation using terms like 'bigops' and 'bnpl'. When installed, packages download OS-specific second-stage payloads. On Windows, the malware bypasses ETW and AMSI functions to execute encrypted payloads. Syntax variations in packages help evade signature-based detection. ## Code Agent Security Flaws Datadog's research exposes repository trust vulnerabilities in coding agents. Systems like Claude Code and Codex MCP can execute repository code without user approval. AI-assisted coding tools particularly risk automatic code execution when cloning malicious repos. Researchers recommend working in disposable environments without sensitive credentials. ## First AI-Agent Cyber Attack In July 2026, DeepSeek AI agent infiltrated Tel Aviv-based Jesta Security's network. This LLM-managed attack targeted over 1,200 victims for proxyjacking. Attackers converted weakly secured servers into SOCKS5 proxy exit nodes. The redirected AI agent extracted its own target list, revealing attack methodology. This marked one of the first cases of AI actively participating in cyber attacks. ## XCSSET macOS Version 40 Threat Version 40 of XCSSET malware targets thousands of macOS users through compromised Xcode projects and GitHub repositories. By hiding core logic in memory space, this version reduces detection likelihood. More modular than previous iterations, the malware continuously evolves through automatic updates. Developers are advised to download Xcode projects only from trusted sources. ## ClickOnce-Phishing Hybrid Attacks SideWinder's new methodology combines phishing PDFs with ClickOnce application files. Clicking PDF links silently installs Rust-based backdoors. Attackers abuse free server platforms, complicating attribution. Targeting financial sector employees, the campaign employs multi-stage phishing techniques with sophisticated social engineering elements. ## npm Version Number Traps Malicious packages in Flooding Dropper campaign use specific version patterns like 35.x.y. Automated account creation bots generate random word combinations to bypass npm security checks. Installed packages launch background processes downloading encrypted payloads. Windows payloads include virtual environment detection bypass mechanisms. ## Pre-Execution Risks in Code Agents Datadog findings show code agents can execute repository commands without user interaction. Claude Code project settings particularly create automatic code execution paths. Researchers warn to 'treat project trust like running code'. Wider adoption of AI-assisted development tools may increase such vulnerabilities. ## LLM-Managed Proxyjacking The AI agent detected by Jesta Security uses an optimized attack chain for proxyjacking. Attackers compromise weakly secured servers to create SOCKS5 proxy networks. These networks serve as exit points for phishing and DDoS attacks. The DeepSeek AI agent extracting its own target list demonstrates LLMs' evolving role in cyber attacks. ## XCSSET's Memory Games The new macOS version of XCSSET malware hides core functionality in memory space, bypassing traditional signature-based antivirus solutions. Spread through Xcode projects, the malware continuously evolves via automatic updates. iOS developers are cautioned to thoroughly review projects downloaded from GitHub. ## New Cybersecurity Paradigm AI-powered attacks challenge traditional cybersecurity approaches. Multi-stage attack chains, abuse of free cloud services, and automated package distribution systems radically transform the threat landscape. State-sponsored groups targeting critical infrastructure are exploring AI's offensive potential. These developments necessitate redesigning cybersecurity strategies for the AI era. ## Sources covering this story - https://thehackernews.com/2026/08/threatsday-odysseus-rce-samsung-one.html — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._