Hello, stay ahead with CISO Brief 🚀

🛡️ Anthropic's coding assistant, Claude Code, can store approved shell commands verbatim in a project file, causing passwords and API keys entered in those commands to be committed to public GitHub repositories. Apiiro found 1,122 public copies of .claude/settings.local.json containing at least one secret, and reproduced the behavior with Opus 5.5 and Fable 5.1. Researchers advise rotating all exposed credentials first, then removing the files from repos and git history and treating agent folders as secret-bearing.

🛡️ Microsoft observed the financially motivated intrusion cluster Storm-3358 using the agentic AI framework ARTEX to target financial organizations in Southeast Asia and globally. The actor combines ARTEX with conventional tools to scan internet-exposed systems, exploit web application flaws, deploy open-source RATs, and abuse compromised credentials and cloud identities for persistence and exfiltration. Microsoft details behavioral signals, detections across its security products, and defensive measures like honeypots and the Security Dashboard for AI.

🚀 Moonshot AI's Kimi K3.1 Pro (Highspeed) is now available on Amazon Bedrock, expanding the open-weight model portfolio with Bedrock's security and governance. This natively multimodal, agentic model offers a 1-million-token context window and a 3-trillion-parameter-class architecture, optimized for high output tokens per second (OTPS). It supports use cases such as coding and document processing, explicit prompt caching to reduce latency and input costs, and runs within the same security boundary as proprietary models.

🔒 This post outlines an architecture for securing unencrypted or unauthenticated protocols (for example, ADB and WebRTC) in multi-tenant Amazon EKS clusters using three reusable components: namespace-scoped mTLS gateways, managed TURN via Amazon Kinesis Video Streams, and a JWT identity chain tied to an enterprise IdP. It assumes cert-manager, AWS Private CA, and Envoy Gateway are available and focuses on design choices and operational learnings rather than step-by-step deployment. The approach reduces attack surface, simplifies TURN operations by removing a custom token-exchange service, and preserves namespace isolation while ensuring auditability of user actions.

🛡️ AWS Agent Registry, available via Amazon Bedrock AgentCore, is now generally available in the AWS GovCloud (US‑West) Region. The service provides a fully managed discovery catalog to publish and organize MCP servers, tools, agents, agent skills, and custom resources, with an approval workflow to control discoverability. Users and AI agents can locate resources via hybrid search, catalog browsing, or a native MCP endpoint. GovCloud availability offers a governed inventory with approval workflows and AWS CloudTrail logging to support public sector compliance and oversight.

🚀 Google Cloud announces general availability of the Storage-optimized Z4M machine series for Compute Engine as both VMs and bare-metal instances. Z4M targets AI/ML training, inference, parallel file systems, and high-throughput storage workloads with up to 168,000 GiB Local SSD and 400 Gbps RDMA networking. Powered by 5th-gen Intel Xeon and Titanium SSDs, Z4M delivers significant IOPS and throughput improvements while supporting Hyperdisk storage options and integration with Cluster Director for large-scale deployments.

🎬 Mito transformed a five‑second generative clip into a 120‑second national TV spot for Movistar by building a serverless, asynchronous pipeline on Google Cloud. They used Cloud Run, Pub/Sub, Cloud Storage, Cloud CDN and Vertex AI to orchestrate over 5,000 assets and maintain real‑time collaboration, low latency, and model consistency. The platform enforces a structured canvas and a "Rule of 6" to control prompt variables while an AI Director copilot translates natural language into model parameters. Continuous internal testing with Gemini and a virtual queuing system helped mitigate API rate limits and ensure production reliability.

🛡️ A researcher found that WordPress sites can be attacked while installing a patch for CVE-2026-119977 because the updater unpacks files using the site's old code into wp-content/upgrade. An attacker who predicts the temporary folder name and reaches the installer before deletion can run code. The update in 7.1.4 randomizes the folder name, but sites should block web access to wp-content/upgrade or perform manual updates outside the web root.

🚀 Amazon MSK Replicator is now available in AWS GovCloud (US‑East, US‑West) to replicate streaming data across Amazon MSK clusters and from external Apache Kafka clusters. MSK Replicator automates asynchronous replication, scaling, and metadata replication (topics, ACLs, consumer offsets) without custom infrastructure. Use the Amazon MSK console or AWS CLI to get started and fail over between regions for resilience.

🔔 Amazon Redshift now permits individual strings up to 16 million bytes within the SUPER data type after you opt in via a database parameter. This applies to ingestion from JSON, PARQUET, TEXT, and CSV sources and during ELT transformations, replacing the prior 65,535-byte limit. The change enables direct ingestion and SQL queryability of large strings, and functions like JSON_PARSE, CONCAT, and LISTAGG can now accept and produce strings up to 16 million bytes. The enhancement is available across all AWS Regions, including AWS GovCloud (US).

🔍 The Residential Proxy Observatory from Layer3 Intel detected proxy exit activity on AS2650, a network registered to the executive office of the US President and operated from the White House. Residential proxies, which route traffic through home or personal devices, mask original IPs and are commonly used by threat actors. Research from Infoblox shows such proxies are widespread across enterprises, and the Observatory also found activity from remote locations like St. Helena.

📊 AWS now amortizes the upfront fee for Amazon EC2 Capacity Blocks for ML across the reservation period in Data Exports for Cost and Usage Report (CUR) 2.0 and AWS Cost Explorer. The amortized cost appears alongside instance usage, showing costs for launched instances and any unused capacity. This enables teams to attribute GPU costs to specific workloads, projects, or applications and identify unused reservation share for better planning. Amortization is available at no extra charge in all Regions except AWS GovCloud (US) and China Regions.

🔒 OpenAI has released GPT-6 Cyber, a security-focused model designed to better find and fix vulnerabilities while remaining aligned to researcher-defined boundaries. The company said the model classifies as Critical under its Preparedness Framework and demonstrates improved adherence to task limits versus GPT-5.6 Cyber. Early results include higher ExploitGym performance and discoveries of numerous Linux privilege escalations and high-severity browser vulnerabilities. Access is initially restricted to select Daybreak Red customers via API and Codex while additional guardrails and misuse detection are enforced.

🔐 This post examines evolving private authentication for the web, focusing on Privacy Pass, blind signatures, and the new MOLE design. It explains how MOLE uses endorsements and anchor sets to hide issuer identity while enabling moderators to issue credentials, and why post‑quantum migration is challenging due to larger primitives. The article highlights VOLE‑in‑the‑Head advances and their role in efficient post‑quantum credential constructions.

🔒 Starting today, you can add Amazon EC2 Capacity Blocks for ML and interruptible Capacity Reservations to Capacity Reservation Resource Groups in AWS GovCloud (US) Regions. This change lets you include On-Demand Capacity Reservations, interruptible reservations, and Capacity Blocks for ML in a single group, simplifying instance launches across reserved capacity. Use launch requests, EC2 Fleet, or Auto Scaling groups to target these groups and set prioritization and fallback to On-Demand. There are no additional charges and the feature is available in all supported AWS Regions except China.

🔍 A new open-source AI-driven offensive framework called ARTEX, released in July 2026, has been adopted by multiple threat actors and appears to be more widespread than initial reports indicated. Censys research found 1,465 publicly exposed ARTEX consoles worldwide, with approximately 815 distinct deployments and many sharing infrastructure with other malicious tools. The original GitHub repo was removed in October after abuse, but forks and commercial variants persist, and several attacks have been linked to the framework since October.

🛠️ Microsoft has published a fix for CVE-2026-83844, a remote code execution flaw in Windows Paint, but the update is delivered only through the Microsoft Store and not via regular Windows updates. The corrected build is 11.2609.401.0; some systems still reported older builds such as 11.2605.81.0 earlier in the week. The vulnerability is an out-of-bounds read rated Important (CVSS 8.1) and requires deep system knowledge to exploit; it was not known to be exploited when disclosed. Administrators should check all per-user and provisioned copies of Paint and update any installs below 11.2609.401.0 via the Store or device-management tooling.

🛡️ Researchers at Forever Security demonstrated that a browser extension able to run scripts on claude.ai could instruct Claude Desktop to execute commands via the Dispatch feature. The flaw requires a user signed in to both claude.ai and Claude Desktop with Dispatch enabled and lets a page-controlled script send prompts and approve prompts shown within the same page. Anthropic told researchers it plans to retire Dispatch by late October, but the functionality remains exploitable and similar features may also be vulnerable.