Hello, stay ahead with CISO Brief 🚀

🛡️ Symantec discovered an offensive agent dubbed ‘Skynet’ that hijacks unprotected AI model endpoints to run autonomous attacks. The agent uses layered jailbreak prompts to override model guardrails, enabling scanning, credential theft, reinfection, and cryptocurrency mining while shifting compute costs to victims. The technique, a form of LLMjacking, exploits misconfigured model servers, test systems, AI gateways, and MCP servers. Symantec captured the attack in a honeypot and advises inventorying model endpoints, enforcing authentication, scoping keys, blocking outbound traffic, and requiring human approval for sensitive actions.

🔒 Amazon announced Route 53 Resolved IP Validation for AWS Network Firewall, launched October 21, 2026, enabling Route 53 Resolver to share DNS answers with Network Firewall. This lets Network Firewall validate outbound connection IPs against the resolver's responses so non-web protocols and TLS SNI can be enforced by domain name. The feature includes a DNS Firewall rule group, a policy setting that checks the first packet, and resolved-domain stateful rules; start in alert mode to observe behavior before enforcing.

🎯 Amazon Elastic Container Service (Amazon ECS) now supports GPU sharing via NVIDIA Multi-Process Service (MPS), allowing multiple ECS tasks to run concurrently on a single NVIDIA GPU. This improves GPU utilization and lowers infrastructure costs for partial-GPU workloads such as small-model inference, embedding generation, and simulation. You specify GPU memory per container and optional compute share; ECS places tasks based on available GPU capacity and supports Managed Instances and ECS on EC2. On Managed Instances, ECS monitors the MPS control daemon and will replace unhealthy instances automatically. The feature is available across all AWS commercial and GovCloud (US) Regions and can be configured using the Console, CLI, SDKs, CloudFormation, or other IaC tools.

🛠️ Amazon SageMaker AI MLflow Apps now support MLflow 3.16, providing a redesigned tracing experience and faster experiment tracking for generative AI and machine learning workflows. The release ships a new default trace explorer with adjustable row height, reorderable columns, and custom columns sourced from any trace tag or metadata field. Metric charts are also accelerated for experiments with large metric histories. New MLflow Apps are created on MLflow 3.16 by default, and existing apps will be upgraded during a maintenance window with no customer action required.

🚀 The VMware solution on Google Cloud is now generally available, offering a self-managed deployment of VMware Cloud Foundation (VCF) on dedicated Google Cloud bare metal. This option complements the Google-managed VMware Engine service by giving customers full administrative control over ESXi, vCenter, NSX, and VCF components while enabling BYOL VCF licensing and integration with existing operational practices. Deployed on GCE Z3 bare metal nodes, the solution provides high-performance NVMe storage, 96 cores, 1.5 TB RAM, and 200 Gbps networking alongside native Google Cloud services.

🚀 Google Cloud and Oracle announced expanded integration at Oracle AI World to deepen Oracle AI Database@Google Cloud, Google’s Agentic Data Cloud, and Gemini Enterprise. The partnership introduces a GA Remote MCP Server with Data API, Knowledge Catalog preview, and a Gemini connector in preview to enable secure, conversational access to Oracle databases. Enhancements include bi-directional Apache Iceberg Lakehouse interoperability, real-time CDC via OCI GoldenGate into BigQuery, BigQuery query federation (preview), expanded regions, and a FedRAMP High pursuit to support regulated workloads.

🛡️ TrendAI and Trend Micro researchers detailed three cyber campaigns in 2026 where AI played central or supportive roles. The incidents ranged from a fully automated CyberStrikeAI operation using 44 agents to human-led attacks aided by large language models and AI-assisted tool development. Targets included VPNs, WebLogic, GIS servers and edge devices across Japan and multiple other countries. The analysis highlights faster weaponization, automated reconnaissance, and commoditized offensive AI tools.

🔒 AWS announced it has completed the Cloud High (PBHH) security assessment with 172 services and features assessed, effective September 18, 2026. The evaluation by the Canadian Centre for Cyber Security (CCCS) makes AWS the first Canadian CSP to meet the Cloud High standard, which builds upon the Cloud Medium profile to address higher integrity and availability requirements. The assessed services span the Canada (Central) and Canada West (Calgary) regions and the summary assessment is available via AWS Artifact.

🔍 This piece profiles three AWS security professionals and the unglamorous daily work that sustains secure systems. It highlights how application and offensive security rely on patient code review, clear writing, empathetic communication, and disciplined practice rather than cinematic exploits. The article emphasizes learnable habits—empathy, judgment, and attention management—and offers practical starting resources like TryHackMe, OWASP, and AWS Skill Builder.

🛡️ Researchers uncovered a vulnerability in OpenAI's API compaction feature that allowed extraction of raw reasoning traces from frontier models. Xint published the technique, called compaction smuggling, on October 26 and reported OpenAI patched the issue on October 23 after confirming it and awarding a $6,000 bounty. The flaw enabled one model's hidden reasoning to be embedded in a compaction and later reconstructed by other models, undermining OpenAI's safeguards against distillation. Xint demonstrated the attack across multiple model versions and raised concerns about large-scale distillation risks.

🧭 This guide explains how to build a secure, governed integration between Google Cloud's Gemini Enterprise and Snowflake using the Model Context Protocol (MCP) and the Google Agent Development Kit (ADK). It outlines an advanced orchestration pattern where a governed Snowflake Cortex Agent generates and executes SQL under the caller's OAuth identity and role. The approach enforces data governance, reduces schema hallucinations, and keeps execution inside Snowflake's trusted boundary.

🚀 Amazon Aurora Serverless for PostgreSQL now supports up to 384 Aurora Capacity Units (ACUs), a 50% increase from the previous 256 ACU limit. Each ACU provides about 2 GiB of memory, enabling up to 768 GiB of memory per serverless database while still scaling down to zero when idle. This improvement targets bursty, agentic AI workloads by providing larger on-demand capacity and the ability to modify existing clusters on Platform Version 4 without downtime.

🛠️ Amazon ElastiCache for Valkey now allows enabling or disabling durability on existing node-based clusters without creating a new cluster. You can choose synchronous writes for cross-AZ persistence with minimal latency or asynchronous writes for microsecond write latency with potential short windows of uncommitted data. The capability expands ElastiCache use cases beyond caching and is available in all Regions supporting Valkey 9.0+ via Console, SDK, or CLI.

🔎 ISC2 has launched the Cyber‑AI Workforce Coalition to help shape the evolving AI security workforce. Announced October 20 and introduced by CEO Scott Beale at ISC2 Security Congress 2026, the coalition will act as a vendor‑neutral convener of employers, standards bodies and practitioners. It aims to produce practical toolkits — not new standards — to help employers write roles, hire and upskill staff for AI security needs. The group’s first meeting is in November, with toolkits expected by spring 2027.

🛡️ Fortinet and Oracle Cloud Infrastructure (OCI) integrate to secure enterprise AI workloads by combining OCI’s managed AI infrastructure with Fortinet’s runtime inspection, governance, and cost-control tools. FortiAIGate acts as a self-hosted, GPU-accelerated gateway that inspects prompts, responses, and agent actions, enforcing guardrails, role-based workspaces, and one-click compliance policies. The joint approach supports centralized controls, data residency choices, and hub-and-spoke network deployments to protect proprietary data and operational workflows.

🤝 Google and Oracle announced a partnership to deploy Google Tensor Processing Units (TPUs) on Oracle Cloud Infrastructure (OCI), with capacity rolling out starting in 2027. The post highlights TPU hardware and system co-design, recent software openness for the PyTorch ecosystem, and the operational benefits of running TPUs inside OCI. Key benefits include data centers built for liquid-cooled AI density, Oracle Acceleron networking integration, proximity to enterprise data, and multi-architecture choice for workloads.

🔔 Amazon announced October 26, 2026 patch updates for multiple Amazon Corretto distributions of OpenJDK. Updated builds — Corretto 27.0.1.0.11, 25.0.5.0.10, 21.0.13.0.11, 17.0.21.0.11, 11.0.33.0.10 and 8u512 — are available for download. Amazon Corretto is a no-cost, multi-platform, production-ready distribution of OpenJDK. Repositories for Apt, Yum, and Apk are provided for Linux systems.

🔒 GoAhead's embedded web server had a critical TLS configuration bug in OpenSSL builds that loaded a certificate revocation list but did not enforce revocation checking, allowing revoked client certificates to authenticate. The vendor, EmbedThis, released GoAhead 6.1.1 on August 2, 2026 (public on October 12) to fix this and two other backend-specific TLS issues. The problems affect builds using OpenSSL, legacy LibreSSL verifiers, or Mbed TLS and require firmware rebuilds by device makers to reach end devices.