🔴 Gitea RCE added to CISA KEV, 🔓 OpenSSL patches 9 CVEs including QUIC double-free, 📦 npm mirror supply chain hijack, 🔑 WordPress TranslatePress unauthenticated takeover, 🗂️ 28K exposed .git repos leak cloud creds, 🏛️ CISA Red Team pwns AD in critical infra

Gitea's official Docker images shipped a default configuration that turned every internet-facing instance into an open door. CISA added CVE-2026-60004 to the Known Exploited Vulnerabilities catalog on August 25 after confirming active exploitation of the remote code execution flaw in the diffpatch Git hook installation mechanism. The vulnerability affects all Gitea versions prior to 1.27.1, and any attacker who can trigger a diffpatch operation gains full server access. Self-hosted Gitea deployments used for internal code review or CI/CD pipelines face particular risk since a single compromised repository can pivot to the broader network. Patches are available in versions 1.27.1 and 1.27.2.

OpenSSL patched nine vulnerabilities in an August 25 advisory spanning versions 1.0.2zr through 4.0.2. The most significant is CVE-2026-18798, a double-free in the QUIC server stack triggered by a malformed INITIAL packet with a Destination Connection ID shorter than eight bytes. The advisory considers remote code execution highly improbable, but the crash is trivial to trigger. A second notable flaw, CVE-2026-63072, produces a deterministic eight-byte out-of-bounds heap write during CMS key unwrapping in CMS_decrypt(). Additional vulnerabilities affect CMP, DTLS, and TLS Raw Public Key code paths. No exploitation in the wild has been reported, but operators running affected OpenSSL versions across QUIC, CMS, or CMP configurations should prioritize patching.

OX Security identified 24 malicious npm packages that exploit package mirror services including unpkg, Yarn, npmmirror, and Tencent to host obfuscated ClickFix phishing pages. The technique is notable because installing these packages does not infect the developer machine. Instead, attackers use npm as free, validated HTTPS storage for a single HTML file, which mirrors then serve directly from their own trusted domains. The phishing pages display fake Cloudflare CAPTCHA verification prompts and redirect victims to credential-harvesting infrastructure including a typosquatted Microsoft domain. The packages typically gained 50 to 300 weekly downloads before removal, and the approach bypasses browser trust warnings because the serving domain is legitimate.

A critical vulnerability in the TranslatePress multilingual WordPress plugin enables unauthenticated account takeover affecting over 400,000 active installations. CVE-2026-19632 (CVSS 9.8) lets attackers extract a raw password-reset URL including the plaintext reset key from the plugin's translation dictionary table via the public trp_get_translations_regular AJAX action, which requires no authentication. The attack chain requires the target administrator's profile to be set to a published secondary language and automatic string saving to be active, which is the default configuration. TranslatePress hooks into wp_mail() to translate outgoing emails, including password-reset messages, and the full reset email with the key gets stored as a translatable string. Wordfence Firewall users were protected before disclosure; the fix is to update beyond version 3.3.1.

Intruder's gitreaper tool scanned 3.5 million internet-facing hosts and found 28,000 publicly accessible .git repositories leaking active credentials for AWS, OpenAI, Stripe, GitHub, and other services. Unlike traditional git tools, gitreaper processes Git objects in memory without downloading full repos to disk. The root cause is a common web server misconfiguration: serving a .git directory over HTTP. Attackers can reconstruct repositories from exposed objects, extract secrets from code history, and gain access to cloud infrastructure, payment platforms, and internal documents. Source code history is often a higher-value target than the current codebase because credentials committed and later removed persist in Git objects indefinitely.

CISA's "A Tale of Two SOCs" advisory published August 25 contrasts red team assessments of two critical-infrastructure organizations, showing how common Active Directory failures turn phishing into enterprise compromise. At the Government Services and Facilities Sector entity, the red team obtained elevated domain privileges, reached sensitive business systems, accessed cloud resources, and eventually read SOC personnel email and deployed keyloggers on defenders' workstations. No effective response interrupted the operation. The assessment exploited the default Machine Account Quota setting and misconfigured AD Certificate Services templates, both of which remain reliable privilege-escalation paths. At the Water and Wastewater Systems Sector facility, defenders contained initial access quickly but still exposed significant identity and cloud security gaps. The core lesson is that large security budgets cannot compensate for a SOC that fails to recognize, escalate, and contain malicious activity.

DigitalOcean's security team detected and blocked an anomalous WordPress admin session attempt on its infrastructure and then reproduced both bypasses for the miniOrange SAML 2.0 Single Sign-On plugin vulnerabilities CVE-2026-61979 and CVE-2026-15981 (CVSS 9.8). The vendor fixed the issues in Standard edition version 17.0.5, but Patchstack's coordination found that the plugin's distribution model, seven plugins with a free version and 30,000 paid customers, obscured actual exposure. CVE-2026-61979 is a signature-algorithm confusion flaw where the plugin accepts HMAC-SHA1 from incoming SAML responses and treats a public RSA key as the HMAC secret. CVE-2026-15981 causes the plugin to interpret an OpenSSL error return as truthy in PHP, letting malformed signatures pass validation.

Group-IB researchers identified Balonx Sistema, a subscription-based PhaaS platform targeting over 20 Mexican financial institutions, harvesting credentials and financial data from more than 1,100 victims since at least October 2025. Unlike standard phishing kits, Balonx maintains a persistent WebSocket connection between the victim's page and the attacker's control panel, enabling real-time monitoring and live page manipulation during active sessions. Operators can push up to 14 fraudulent screens. The platform also deploys Android malware for mobile credential theft and uses AI-powered voice fraud to supplement the phishing chain. Payments are automated through cryptocurrency exchange services, structuring the operation like a commercial SaaS business with Individual and Office subscription tiers.

Core Werewolf, a cluster previously reliant on legitimate UltraVNC remote-access tools, has deployed its first full-featured C++ RAT called CoreRAT targeting Russian public-sector and defense-industry organizations. Researchers at bi.zone identified two delivery methods: 7z self-extracting archive droppers and Rust-based droppers, both paired with a PDF decoy written in Russian to resemble official military or government correspondence. CoreRAT provides reconnaissance, command execution, file transfer, network discovery, and self-deletion capabilities. The shift from abusing legitimate remote-administration software to developing a purpose-built RAT represents a meaningful step up in the group's malware-development capability and may help evade detections tuned for known tools.

ClearFake operators have deployed WordlistLoader, a new loader that encodes shellcode as sequences of ordinary English words where each word maps to a single byte based on its position in a build-specific 256-word list. The loader reads each word, finds its index, and writes the corresponding byte into memory, reconstructing the payload at runtime. This makes static analysis significantly harder since the encoded content looks like a harmless word list on disk. The delivery chain uses compromised legitimate websites displaying fake CAPTCHA overlays, and the loader includes anti-analysis features that scan loaded Windows modules and restore original bytes if it detects security-product hook instructions, effectively unhooking AV or EDR monitoring before the Amatera Stealer payload executes.

A multi-agent AI framework built on Hermes and OpenClaw autonomously compromised 85 government accounts in Asia over approximately four days in July 2026, exfiltrating at least 2,564 personnel records. The framework deployed up to eight letter-designated sub-agents across 12 attack waves, decompiled Angular JavaScript bundles, enumerated 21 connected systems, mapped OIDC and Keycloak configurations, and identified over 36 API endpoints on a single target. It paired harvested usernames with predictable password patterns and used Tesseract OCR to defeat CAPTCHA challenges, initially cracking 12 accounts before expanding its pattern set to compromise 73 more. Three exposed debug-style authentication endpoints issued valid sessions, and a JWT validation weakness using the "none" algorithm allowed further lateral movement.

PavinLoader is a multi-stage .NET malware loader observed across at least three distinct delivery campaigns including ClickFix social engineering lures, fake software installers, and malicious RenPy game packages. The loader abuses MSBuild executables with project files that use UsingTask and assembly-loading functions to execute malicious .NET DLLs, which are trojanized versions of legitimate open-source libraries including DotNetZip, Nancy, Renci.SshNet, and OpenXML. The final payload is Amatera Stealer distributed via blockchain-based C2 infrastructure. The heavy obfuscation and legitimate-tool abuse make detection harder for endpoint security that trusts signed Microsoft binaries, and the loader's appearance across multiple delivery campaigns suggests it may operate under a Loader-as-a-Service model.

Broadcom's August 20 advisory disclosed 91 CVEs affecting Spring Framework and related projects, with Sonatype estimating 209,569 downstream software components affected. The standout vulnerability is CVE-2026-59285, a Spring for GraphQL deserialization flaw with a CVSS of 9.2 that enables remote code execution in applications using Spring for GraphQL with Jackson 2.x JSON deserialization, paginated GraphQL fields, and exploitable classes on the classpath. The volume signals continued pressure on the Java/Spring ecosystem as AI-accelerated fuzzing tools surface vulnerabilities faster than most organizations can patch. GraphQL endpoints are typically internet-facing and process structured user input, raising the attack surface significantly.

Palo Alto Networks patched multiple vulnerabilities in the GlobalProtect App that allow local privilege escalation to SYSTEM on Windows or root on Linux and macOS, affecting GlobalProtect App versions 6.0, 6.2, and 6.3. Local privilege escalation in a VPN client matters because the client runs with elevated trust on endpoints that are often the first hop into a corporate network. An attacker who already has code execution as an unprivileged user can use these flaws to take full control of the host, turning a modest foothold into a platform for credential theft, lateral movement, and persistence. Managed fleets should track GlobalProtect App version inventory and plan the upgrade, since client-side fixes depend on users or MDM pushing the patched build.

Watchlist: update all Gitea deployments to version 1.27.2 or later given the KEV addition and confirmed exploitation of CVE-2026-60004; patch OpenSSL across QUIC, CMS, and CMP configurations for the nine newly disclosed vulnerabilities; update TranslatePress beyond version 3.3.1 on all WordPress sites and audit translation dictionary tables for evidence of password-reset URL extraction; block HTTP access to .git directories at the web server or CDN level and rotate any credentials found in exposed repositories; audit Active Directory Machine Account Quota settings and AD Certificate Services template configurations following CISA's red team findings; update miniOrange SAML SSO WordPress plugin to version 17.0.5 and review authentication logs for suspicious admin sessions; and monitor for Balonx Sistema infrastructure targeting Mexican financial institutions and CoreRAT delivery chains using Russian-language PDF decoys.