AWS Security Enhancements, GitHub Outage, and Critical Patches

Cloud platforms rolled out security-focused capabilities while researchers and vendors disclosed critical vulnerabilities and ongoing threat activity. AWS announced multiple enhancements spanning firewall telemetry, AI model access, analytics, and operations. At the same time, organizations faced a major GitHub outage and fresh exposure claims, and defenders received new patch guidance for high-impact flaws across developer platforms, WordPress plugins, and mobile chipsets.

AWS Expands Security Telemetry, Resilience, and AI Options

Amazon Bedrock added support for OpenAI’s GPT‑5.6 family (Sol, Terra, Luna) through the bedrock-runtime endpoint and extended compatibility to Responses, Converse, and Chat Completions APIs. The release also introduced cross‑Region inference with Geo and Global routing modes. Geo confines processing to a defined geography (including a new US Geo), while Global serves requests from any commercial AWS Region offering the model and can provide lower per‑token pricing compared with in‑Region and Geo modes. The models inherit Bedrock’s account‑level controls and observability, including S3 or CloudWatch Logs delivery for invocation logs and CloudWatch metrics for token usage and latency; spend can be itemized via Cost Explorer and the Cost and Usage Report. Separately, OpenSearch Service enabled automatic semantic enrichment for VPC domains, bringing AI-driven, context-aware search to private deployments on OpenSearch version 2.19 or later across 11 Regions without requiring customers to manage ML models.

AWS Network Firewall now provides stateful rule hit counts, delivering per-rule visibility to help validate policies, identify obsolete or shadowed rules, and speed incident response. Metrics can refresh as frequently as every five minutes and are enabled by default for custom and managed rule groups at no additional charge (standard log storage and query costs still apply); availability excludes the Middle East (UAE) and Middle East (Bahrain) Regions. For capacity management, EC2 Auto Scaling added batch instance termination, allowing up to 100 instance IDs in a single API call. The service validates the batch atomically before proceeding and preserves lifecycle hooks and load balancer draining, reducing API overhead and accelerating scale-downs for large, transient fleets without extra cost.

Amazon MSK introduced cluster-level custom domain names for Provisioned clusters in both ZooKeeper and KRaft modes. Administrators can define a domain once and have it automatically applied and persisted across scaling, migrations, and disaster recovery, simplifying client connectivity, naming consistency, and operational continuity at no extra charge. In productivity tooling, Amazon Quick extensions for Microsoft 365 became generally available in multiple Regions. Quick integrates into Excel, PowerPoint, Word, and Outlook to automate spreadsheet analysis, generate branded decks, produce and review documents with track changes, and manage inbox and calendaring using organization-defined templates and context.

Critical Vulnerabilities and Patch Guidance

GitLab released out-of-schedule updates for a critical GraphQL vulnerability (CVE‑2026‑19478, CVSS 9.4) that could allow unauthenticated attackers to modify or delete public projects and user data on self-managed instances. Fixes are available in versions 19.2.4, 19.1.6, 19.0.8, and 18.11.11; GitLab.com and GitLab Dedicated are already patched. A second issue, CVE‑2026‑19650 (CVSS 7.1), addresses a CSRF weakness in GraphQL multiplex query handling requiring user interaction. GitLab plans to publish technical reports approximately 90 days after release.

Forminator plugin for WordPress received a fix (version 1.56.2) for a critical arbitrary file upload vulnerability (CVE‑2026‑15748, CVSS 9.8) that could enable unauthenticated remote code execution via malicious PHP uploads under certain storage configurations. Wordfence also highlighted a separate critical authentication bypass in User Profile Builder (CVE‑2026‑15826, CVSS 9.8), patched in 3.16.5, which could allow an attacker to obtain administrator sessions when specific site settings are enabled. Administrators are advised to update both plugins and review upload storage and registration configurations.

Microsoft Defender is affected by a newly disclosed local privilege escalation dubbed ShieldBreak (CVE‑2026‑69414). A publicly released proof-of-concept reportedly achieves SYSTEM privileges on fully patched Windows 10, Windows 11, and Windows Server when Defender is enabled. Microsoft is investigating and preparing a security update; organizations should monitor advisories and plan to deploy the fix when available.

Unisoc exploit chain research detailed a two-stage path to Android kernel execution via modem firmware vulnerabilities. An attacker controlling a private 4G (VoLTE) network and a victim who answers a video call could enable modem-level code execution and then disable Memory Protection Unit safeguards to access kernel memory. Multiple Unisoc chipsets are affected, no CVE is assigned, and no vendor patch was available at publication.

Snowflake repo maintainers fixed a GitHub Actions workflow injection in the public snowflakedb/snowflake-connector-net repository that allowed command injection via crafted GitHub issues. During authorized testing, researchers exfiltrated a Jira API token from the Actions runner; Snowflake merged a fix the same day and rotated the token the following day, reporting no evidence of unauthorized access during a five-day exposure window.

Operational Incidents and Exposure Claims

GitHub experienced a widespread outage impacting core services, with elevated error rates across web and API traffic, significantly higher failure rates for archive and raw downloads, and degraded performance for Actions, Webhooks, Issues, Pull Requests, and authentication systems. Copilot availability was also reduced. Some components remained operational, and the root cause has not yet been disclosed.

Clop claims of data theft against General Electric and Philips are under investigation by the companies. Philips reported containing an attempted compromise of an internal enterprise server without customer impact; GE is assessing the claim. The listings tie to exploitation of a critical PTC Windchill/FlexPLM flaw (CVE‑2026‑12569) for which patches were released in June; agencies and researchers reported webshell deployment and data exfiltration from affected instances.

Azure records exposure claims by a threat actor marketing 3.64 million alleged employee directory entries from multiple major companies remain unverified. Sample data reportedly includes names, emails, job titles, phone numbers, addresses, and service and admin account identifiers. Some contacted organizations stated the data appears dated and limited in sensitivity, but such details can aid social engineering and targeted attacks.

Weekly recap reporting highlighted rapid exploitation of recent vulnerabilities, use of exposed services and existing access for high-impact incidents, and techniques targeting browser sessions and supply-chain tooling. Recommended actions include prioritizing patches for listed CVEs, reviewing remote-access exposure, strengthening browser and session protections, and tightening governance around emerging AI toolchains.

Evolving Tradecraft and Malware

Cavern C2 continues to evolve with modular components, DNS-based control, and optional relaying via Google Apps Script to blend into legitimate traffic. Researchers also documented a broker component for dynamic module loading and a separate tool, HOLLOWGRAPH, that uses Microsoft Graph calendar events as covert dead-drops and DNS tunneling to refresh credentials, complicating network-based detection.

AmnesiaStealer is a multi-stage macOS infostealer that uses a fake GitHub download page and Terminal lure to deploy a Rust payload. Beyond credential and data theft, a separate module clones Chromium profiles and establishes a WebSocket backchannel to remotely drive the victim’s browser via DevTools protocols, effectively turning authenticated sessions into attacker-controlled browsing.

Zhipu GLM‑5.3, a coding-focused AI model, reportedly achieved strong internal benchmark results for vulnerability identification following post-training scaling and reinforcement learning in longer-horizon environments. Testing with security teams surfaced thousands of issues across hundreds of projects, and the company plans to release model weights after safety evaluation, underscoring concerns about dual-use capabilities of advanced coding AIs.