< ciso
brief />
Autopilot Expands, Post-Quantum WARP Ships, Cisco Zero-Day Patched

Autopilot Expands, Post-Quantum WARP Ships, Cisco Zero-Day Patched

Coverage: 24 Sept 2025 (UTC)

< view all daily briefs >

Cloud and AI platform announcements led activity on 24 September 2025. Google extended GKE Autopilot capabilities to Standard clusters, and Microsoft and Google detailed architectures for agentic AI. Cloudflare published several post-quantum encryption and web content initiatives, while U.S. prosecutors tied an alleged Scattered Spider member to at least $115 million in ransoms. Cisco patched an actively exploited IOS and IOS XE zero-day, and researchers described long-running espionage campaigns against edge appliances and virtualization infrastructure.

Cloud Platforms Expand Autopilot Compute and Agentic AI Tooling

Google Cloud has made key components of GKE Autopilot generally available to qualifying Standard Google Kubernetes Engine clusters, extending access beyond dedicated Autopilot clusters. The initial rollout covers clusters in the Rapid release channel running GKE version 1.33.1-gke.1107000 or later. Clusters in the Extended channel and those using legacy routes-based networking are excluded until later updates. The functionality is delivered through two built-in compute classes, autopilot and autopilot-spot. Teams can assign them to individual Pods or set them as a namespace default, which removes the need to plan node shapes and sizes for most workloads. The container-optimized compute platform supports pods from 50 milli-CPU up to 28 vCPU, bills on pod resource requests rather than unused node capacity, and supports both horizontal and vertical autoscaling. Workloads that need more than 28 vCPU or specific hardware, including GPUs and TPUs, can run on specialized compute with node-based pricing through customized compute classes. A new compute-class provisioning mode creates node pools per compute class automatically without changing scheduling behavior for existing node pools, which allows gradual adoption. Google reports that about 30% of active GKE clusters were created in Autopilot mode in 2024.

Microsoft closed its Agent Factory series with a proposal for an open agentic web stack, authored by Yina Arenas and Ulrich Homann. The post lists eight components intended to make multi-agent ecosystems secure, scalable and interoperable across organizational boundaries:

a communication protocol service, a discovery registry, identity and trust management, tool invocation and integration, orchestration, telemetry and observability, memory, and evaluation and governance.

It points to open standards such as MCP (Model Context Protocol) and A2A (Agent-to-Agent) for messaging and tool invocation. Identity, using OIDC, JWT and Microsoft Entra ID integrations, is positioned as the control plane for authorization, accountability and zero-trust posture. Azure AI Foundry is presented as a concrete implementation. It unifies frameworks including Semantic Kernel, AutoGen and LangGraph, embeds MCP for vendor-neutral tool access and extends OpenTelemetry for agent-aware tracing. The authors recommend that organizations adopt open standards, invest in identity, observability and memory, and embed governance policies directly into workflows. Microsoft plans further demonstrations at Ignite 2025.

Google used Big Data London to present an AI-native data stack aimed at turning data scientists into builders and operators of autonomous agents. Colab Enterprise notebooks now include native SQL cells and an enhanced Data Science Agent that can plan and execute toolchains across BigQuery ML, DataFrames and Spark. The Lightning Engine for accelerating Spark workloads is generally available. Spark 4.0 support and stateful processing for BigQuery continuous queries are both in preview. A "Build-Deploy-Connect" toolkit pairs the Agent Development Kit with first-party BigQuery connectors and the MCP Toolbox for secure data access. Part of that toolkit, the Gemini CLI extensions for Google Data Cloud bring Gemini capabilities into developer terminals. They cover BigQuery, AlloyDB, Cloud SQL for MySQL, PostgreSQL and SQL Server, Firestore, Looker and Spanner. The extensions can generate and run natural-language queries, create and inspect Cloud SQL instances, and provision database users, including Cloud IAM users. Administrators must enable the required APIs, configure IAM permissions and set extension-specific environment variables, so the tools rely on standard cloud configuration and security controls.

OpenAI is internally testing GPT-Alpha, an agent build that the company indicates uses GPT-5 for reasoning and tool use. An unreleased feature was briefly exposed to users in an accidental push. A screenshot shared on X by Clay Malott showed an option labeled "Agent with Truncation" under Alpha Models. The agent's internal prompt lists the following capabilities:

browsing the web, generating and editing images, writing, running and debugging code, and creating or editing documents, spreadsheets and slides.

The prompt also instructs the agent not to access or disclose private information unless the user supplies it. No public rollout timeline has been given. Sam Altman has previously said advanced features are likely to reach paying customers first because of compute costs. The combination of live web access and code execution expands the attack surface and raises data-leakage concerns, placing emphasis on access controls, provenance and audit logging.

AWS announced two infrastructure changes. Lambda code signing is now enforced in AWS GovCloud (US-West and US-East) using AWS Signer. Administrators create Signing Profiles and restrict functions to accept only specific profiles. They can also choose whether signature failures emit warnings or block deployments. Access to these operations is governed through IAM, and AWS states there is no additional charge. Separately, AWS removed network burst limits for EC2 I7i and I8g instances larger than 4xlarge. For those sizes, the change doubles always-available network bandwidth and eliminates the network I/O credit mechanism. The families offer up to 1.5 TiB of memory, 45 TB of local storage, up to 100 Gbps of network performance and 60 Gbps of dedicated EBS bandwidth. Sizes at or below 4xlarge retain baseline-plus-burst behavior.

Cloudflare Advances Post-Quantum Encryption and Crawler Controls

Cloudflare's WARP client now supports post-quantum key agreement to protect tunneled traffic against harvest-now-decrypt-later threats. The client covers the consumer 1.1.1.1 app and the enterprise Cloudflare One Agent. WARP carries traffic in MASQUE tunnels to Cloudflare's network and can deliver it to origins over post-quantum Cloudflare Tunnels where the origin supports PQC. To meet FIPS and FedRAMP requirements, Cloudflare currently uses a hybrid that pairs P-256 ECDH with an early ML-KEM variant (P256Kyber768Draft00). Post-quantum signatures and certificates will follow once standardization matures. The rollout uses an API flag with three states (No PQC, PQC with allowed downgrades, and PQC-only), plus an MDM override that lets administrators enable PQC-only mode immediately. Supporting versions are already available: desktop 2025.5.893.0, iOS 1.11 and Android 2.4.2. The consumer Phase 1 rollout is complete, and Cloudflare One Agent Phase 1 is expected by the end of 2025. PQC-only mode is planned for mid-2026.

Cloudflare also detailed Automatic SSL/TLS, which has upgraded origin-facing encryption for more than 6 million domains. Over 4 million of those domains moved from plaintext or Flexible modes to Full or Full (Strict). The system starts a recommended mode at 1% of traffic, ramps it to 100%, and rolls back immediately on TLS failures. Cloudflare reports a 7–8% reduction in plaintext origin connections, with TLS 1.3 now accounting for nearly 60% of encrypted origin traffic. It plans to begin automatic post-quantum handshake testing in Q4 2025 using hybrid X25519 and ML-KEM key agreement.

Cloudflare is also prototyping a SaaS-to-SaaS proxy in response to marketplace integration breaches such as the Salesloft incident. Two deployment models are under development:

a customer-controlled front door that uses vanity or branded hostnames, and a vendor-side reverse proxy that extends API Shield capabilities to SaaS platforms.

The design centers on key splitting, which divides bearer tokens into two unusable fragments. The integration holds one fragment, and Cloudflare's edge stores the other transiently, so complete tokens are never persisted. A token is reconstructed in memory only for the duration of a forwarded call, and deleting the edge fragment can revoke access in seconds. Planned additions include token and session logging, session anomaly detection and token replacement. The approach requires marketplace apps to support vanity or branded domains for OAuth flows.

On automated access to web content, Cloudflare proposed five responsible AI bot principles:

1. Public disclosure of bot identity, operator and purpose.

2. Truthful self-identification, with a push toward cryptographic verification through Web Bot Auth.

3. A single declared purpose per crawler.

4. Respect for webmaster preferences expressed through robots.txt and evolving HTTP headers.

5. Good intent, meaning bots avoid excessive load, deception and stealth.

The company argues that user-agent strings and IP lists are insufficient because they can be spoofed. Its new Content Signals Policy extends robots.txt with three machine-readable signals: search, ai-input and ai-train, each set to yes or no. Cloudflare will update managed robots.txt files for roughly 3.8 million domains to include "Content-Signal: search=yes, ai-train=no", with an opt-out in the dashboard. Cloudflare describes the signals as expressions of preference and legal reservation rather than technical countermeasures, referencing Article 4 of EU Directive 2019/790. It recommends pairing them with WAF and Bot Management. The specification is released under a CC0 license.

Ransomware Charges and a Collapse Traced to One Password

U.S. prosecutors unsealed charges against 19-year-old Thalha Jubair, alleging a central role in Scattered Spider and linking the group to at least $115 million in extortion payments. Jubair and 18-year-old Owen Flowers appeared in a London court this month. Prosecutors trace more than 120 intrusions against 47 U.S. entities to the group, from 2021 through September 2025. The activity spans SIM-swapping, mass SMS phishing, intrusions at business process outsourcing firms, and ransomware incidents. Authorities seized a server tied to Jubair through which proceeds allegedly flowed, and $36 million in cryptocurrency was specifically seized. The complaint connects Jubair's aliases to earlier LAPSUS$ campaigns and to Telegram channels such as Star Chat that marketed SIM-swap and phishing services. U.S. charges include computer fraud, wire fraud and money laundering, with a theoretical maximum of 95 years. U.K. prosecutors separately charged both defendants under the Computer Misuse Act.

A separate account details how KNP Logistics Group collapsed in June 2025 after the Akira ransomware group guessed a single employee password. KNP was a 158-year-old Northamptonshire transport company operating roughly 500 trucks. The targeted internet-facing account lacked multi-factor authentication. The attackers deployed ransomware across the environment and destroyed backups and disaster-recovery systems. The gang demanded an estimated £5 million. Despite industry-standard IT compliance and cyber insurance, the company entered administration, and about 700 jobs were lost. The report cites Kaspersky analysis finding that 45% of 193 million compromised passwords could be cracked within a minute. Its recommendations include strong password policies, breached-password detection, enforced MFA on internet-facing accounts, least-privilege access, and isolated, regularly tested backups.

Exploited Network Flaws and Espionage Against Edge Appliances

Cisco has released patches for CVE-2025-20352, a high-severity zero-day in IOS and IOS XE that has been exploited in the wild. The stack-based buffer overflow sits in the SNMP subsystem and affects any device with SNMP enabled. A crafted SNMP packet over IPv4 or IPv6 can trigger it. Authenticated attackers with low privileges can cause denial of service, while attackers with elevated privileges can execute arbitrary code as root on IOS XE systems. Cisco PSIRT became aware of exploitation after local Administrator credentials were compromised. There are no complete workarounds, and Cisco recommends upgrading to the fixed releases. Where that is not immediately possible, administrators should restrict SNMP access to trusted users and interfaces. The same update cycle addressed 13 further vulnerabilities. Two of them, CVE-2025-20240 and CVE-2025-20149, have proof-of-concept code.

Firmware security firm Binarly reported Supermicro BMC flaws that allow attackers to flash maliciously crafted firmware images that persist across reboots. The findings include three issues:

a bypass of the previously patched CVE-2024-10237, a new Root of Trust validation bypass tracked as CVE-2025-6198, and a related issue tracked as CVE-2025-7937.

Together, these issues allow signed regions to be injected or relocated while the image digest and signature remain valid. Researchers also showed that kernel authentication is not enforced during boot on affected firmware. Binarly published proof-of-concept exploits, and Supermicro has released September 2025 firmware updates for impacted models.

Google Threat Intelligence Group and Mandiant detailed the BRICKSTORM campaign, attributed to the China-nexus cluster UNC5221. The campaign targets edge and management appliances in the legal services, SaaS, BPO and technology sectors. BRICKSTORM is a Go backdoor with SOCKS proxy functionality, designed to run on appliances that lack EDR visibility. The median dwell time is about 393 days. After gaining a foothold, the actor:

moves laterally to VMware vCenter and ESXi using captured credentials; deploys a credential-stealing Java Servlet filter tracked as BRICKSTEAL; clones sensitive VMs such as domain controllers to extract files including ntds.dit; and installs a JSP web shell tracked as SLAYSTYLE.

Google provides YARA rules, a scanner and prioritized hunts. Its hardening recommendations include inventorying appliances, forwarding logs to a SIEM, enabling vSphere lockdown mode and MFA, treating credential vaults as Tier 0, and restricting appliance egress to vendor endpoints.

Recorded Future named a separate likely Chinese state-sponsored actor, RedNovember. The group was previously tracked as TAG-100 and overlaps with Microsoft's Storm-2077. Between June 2024 and July 2025, it exploited internet-facing perimeter appliances. Targets included Check Point (CVE-2024-24919) and Palo Alto Networks (CVE-2024-3400) products, as well as equipment from Cisco, Citrix, F5, Fortinet, Ivanti and SonicWall. Tooling included the Go-based Pantegana framework, Spark RAT, a LESLIELOADER variant and Cobalt Strike. Victims included ministries of foreign affairs, defense and aerospace entities, at least two U.S. defense contractors and a Southeast Asian trade-focused intergovernmental body. Activity concentrated on Panama, the U.S., Taiwan and South Korea.

Trend Micro researchers disclosed two critical flaws in Wondershare RepairIt: CVE-2025-10643 (CVSS 9.1) and CVE-2025-10644 (CVSS 9.4). The flaws stem from overly permissive storage tokens embedded in the application code and from unencrypted cloud buckets. Exploitation could expose user-uploaded images and videos, AI models, binaries, container images and source code. Because the application retrieves and executes models from that storage, tampered artifacts could reach customers, creating a supply chain risk. Trend Micro reported the issues through the Zero Day Initiative in April 2025 but has received no vendor engagement. With no patch available, it recommends restricting interaction with the product.