Control Planes Advance As AI Phishing and Patch Urgency Rise

A new management layer for enterprise control landed in public beta from Cloudflare, while federal agencies were ordered to remediate a Fortinet EMS flaw under active exploitation, as reported by BleepingComputer. The day balanced platform consolidation and governed analytics with advisories on AI‑enabled phishing, fast‑moving ransomware operations, and emerging hardware exploitation techniques.

Control planes tighten

The public‑beta launch of Organizations adds a unified layer above multiple accounts so large enterprises can centralize oversight while preserving least‑privilege boundaries. The model introduces an Org Super Administrator with cross‑account authority that does not require membership in each account, plus shared policy sets for WAF and Gateway rules, a flat account list, and roll‑up HTTP analytics. Behind the scenes, a substantial refactor improved permission‑enumeration performance and removed legacy code; the beta starts with enterprise customers and a roadmap that includes organization‑level audit logs, billing reports, and additional roles. The approach aims to reduce operational sprawl without weakening per‑account isolation.

On the security platform side, Fortinet detailed FortiOS 8.0 as its unifying operating system across on‑prem, hybrid, multi‑cloud, and IT/OT environments. New capabilities focus on four emerging risks: visibility into autonomous agent activity via MCP observability; image‑based data exfiltration countered with OCR in FortiGuard DLP; alert overload reduced through embedded agentic automation; and long‑term cryptographic exposure mitigated with hybrid cryptography aligned to FIPS 204/205, extending to SSL deep inspection. Why it matters: consolidating policy and telemetry under a single fabric helps cut complexity while addressing AI‑driven behaviors and future‑proofing for post‑quantum transitions.

CrowdStrike introduced Continuous Visibility in Falcon Exposure Management to answer a recurring question faster: “Does this new CVE affect our assets?” By retaining network asset metadata and separating data collection from cloud‑driven evaluation, the platform re-evaluates stored data automatically when new detection content ships, surfacing exposures between scans and enabling one‑click targeted rescans. The design aims to prioritize remediation without raising operational overhead or scan frequency, matching a threat tempo where eCrime breakout times continue to shrink.

Governed analytics, now conversational

Self‑service Explores are now available in Looker, letting users drag‑and‑drop local files or import Google Sheets to create ad‑hoc Explores stored in the customer’s BigQuery environment. Merge queries can enrich uploads with modeled data, and conversational analytics supports natural‑language exploration and drill‑downs, while administrative controls distinguish experimental datasets from governed models. The capability speeds hypothesis testing and temporary joins without bypassing existing governance.

To extend these experiences into products, Looker Embedded added a generally available Conversational Analytics API that brings multi‑turn, Gemini‑powered querying, AI recommendations, a code interpreter, and verifiable SQL into embedded workflows via iframe or SDKs. By leaning on the governed semantic layer, the feature aims to reduce hallucinations and improve auditability for integrated BI and AI agent experiences.

Edge, developer, and infrastructure shifts

AWS released a compact Greengrass component SDK with native bindings for C, C++, and Rust that cuts component footprints to under 0.5MB—down from roughly 30MB—opening deployment on highly constrained edge devices. In parallel, AWS announced GA of Smithy‑Java, an open‑source framework that generates type‑safe Java clients and standalone types from Smithy models, using Java 21 virtual threads for a simpler blocking‑style API and schema‑driven serialization to shrink SDK size and latency. Together, these updates target performance‑ and cost‑sensitive embedded use cases and faster service delivery in Smithy ecosystems.

Google Cloud reported that its seventh‑generation Ironwood TPU delivers an approximate 3.7× improvement in compute carbon intensity over TPU v5p based on January 2026 fleet measurements, alongside about 5× more utilized FLOPs. Gains come from hardware, sparse architectures like MoE, FP8 precision, and orchestration that raises utilization. For AI workloads, lower CO2e per unit compute depends on model sparsity, precision choices, and real‑world utilization.

Active operations, exploits, and policy pressure

The Defender research team at Microsoft detailed a large AI‑enabled phishing campaign abusing OAuth Device Code Authentication. An automated pipeline used generative lures and thousands of short‑lived nodes on serverless/PaaS platforms to generate device codes in real time when victims clicked, enabling seamless session takeover even with MFA. Stolen tokens supported Microsoft Graph reconnaissance, device registration for long‑lived tokens, malicious inbox rules, and targeted exfiltration; mitigations include restricting or blocking device code flow, enforcing phishing‑resistant MFA, Conditional Access, and revoking refresh tokens. Separately, Microsoft attributed high‑tempo Medusa ransomware operations to Storm‑1175, which rapidly weaponizes N‑days and has used multiple zero‑days to breach web‑facing systems before moving to data theft and encryption—often within days and sometimes within 24 hours—while leveraging legitimate RMM tools, credential dumping, and Defender tampering.

Exploit and research activity spanned platforms. A researcher publicly released “BlueHammer,” a Windows local privilege‑escalation exploit chaining TOCTOU and path confusion to access the SAM database, as covered by BleepingComputer; the PoC is not trivial and lacks a patch at time of reporting. In parallel, BleepingComputer reported on GPUBreach, which Rowhammers GDDR6 to corrupt GPU page tables for arbitrary GPU memory access and chains into CPU‑side escalation via driver bugs, demonstrated on an NVIDIA RTX A6000; mitigations remain limited, with ECC insufficient against multi‑bit flips.

Cryptocurrency targeting and cloud abuse persisted. BleepingComputer described a theft exceeding $280 million from Solana‑based Drift Protocol after attackers seized Security Council powers following a months‑long, in‑person social‑engineering buildup and suspected developer compromise. In Kubernetes environments, Unit 42 observed a 282% YoY rise in operations, with case studies showing RCE in containers followed by service‑account token theft, RBAC abuse, cloud credential exfiltration, and cryptomining—reinforcing the need for audit logs, runtime telemetry, and strict least‑privilege. Regionally, The Hacker News relayed Check Point’s attribution of an Iran‑linked password‑spraying campaign against 300+ Microsoft 365 tenants in Israel and others in the U.A.E., with exfiltration attempts and overlaps to prior Iran‑nexus tooling; enforcing MFA and Conditional Access remains central. Supply‑chain risk also resurfaced as The Hacker News detailed how backdoored LiteLLM PyPI packages harvested credentials across developer machines and transitive dependencies—prompting calls for secrets governance, rotation, and developer endpoint hardening.

Separately, a policy case in New Mexico advanced a theory that privacy‑protective end‑to‑end encryption can constitute a harmful “design choice,” with potential remedies to weaken it. As Schneier summarized, critics warn such measures would chill deployment of security features and risk broader exposure to surveillance, breaches, and abuse.