AI Platforms Harden, Managed Defense Expands, Exploits Surge
Coverage: 17 Feb 2026 (UTC)
< view all daily briefs >Bedrock added new model and customization options while managed detection matured at the SOC layer. Amazon’s platform now offers Claude Sonnet 4.6, and adds managed reinforcement fine-tuning for open‑weight models, and Unit 42 unveiled a 24/7 managed XSIAM service built for machine‑speed response. Together, these moves underscore a shift toward stronger defaults, operational guardrails, and expert‑operated defense for AI‑heavy environments.
AI platforms and device safeguards advance
Amazon expanded Bedrock’s model roster as Claude Sonnet 4.6 becomes available, emphasizing improved multi‑step orchestration, context compaction, and consistent conversational quality at lower cost than Opus 4.6. The release targets domain‑specific work such as coding support, compliance reviews, and data summarization, with minimal migration friction from Sonnet 4.5, and immediate availability across supported regions via the Bedrock console (AWS). In parallel, Bedrock introduced managed reinforcement fine‑tuning for popular open‑weight models and OpenAI‑compatible APIs, enabling rule‑based or AI‑judge reward functions, small‑data iterative training, and direct deployment to inference endpoints without extra steps; proprietary data remains within AWS’s governed environment (AWS).
On mobile platforms, Apple’s iOS and iPadOS 26.4 developer beta brings end‑to‑end encryption for RCS messaging between compatible Apple devices and expands Memory Integrity Enforcement, allowing apps to opt into always‑on memory safety beyond prior soft mode. The beta also moves Stolen Device Protection toward default enablement, with biometric requirements in sensitive flows and a delay before password changes (Infosecurity). Google previewed Android 17 changes that raise the baseline: cleartext traffic is blocked by default unless explicitly allowed via network security configuration, HPKE support is exposed via a new SPI, certificate transparency is on by default, and new install‑time permissions tighten localhost interactions. Runtime improvements aim to cut CPU overhead and smooth frames, accompanied by a Canary channel for earlier testing (Infosecurity).
Managed defense and analytics
Palo Alto Networks introduced Unit 42 Managed XSIAM 2.0, a 24/7 SOC service operated by analysts, hunters, and responders on Cortex XSIAM to compress detection‑to‑response cycles. The service continuously engineers detections, correlation, and automated playbooks from frontline intelligence, supports native and third‑party EDR telemetry, and can execute pre‑authorized, expert‑validated remediation across endpoints, firewalls, identity, and cloud. A breach response guarantee and outcomes such as material reductions in response time are highlighted for organizations pursuing consolidation and machine‑speed operations (link in lead).
In data analytics, Google Cloud previewed global queries for BigQuery, letting teams join and aggregate datasets across regions with a single SQL statement. BigQuery orchestrates region‑specific sub‑queries and moves optimized partial results, while honoring security controls (for example, VPC Service Controls). The feature is opt‑in and permission‑gated, aiming to minimize cross‑region transfers and associated costs while shortening time‑to‑insight (Google Cloud).
Palo Alto also framed the rise of the agentic endpoint—extensions, plugins, MCP/local servers, containers, and model artifacts installed outside traditional controls—as a growing blind spot. The company announced its intent to acquire Koi to enhance visibility, risk understanding, and real‑time policy enforcement across AI‑native ecosystems, and cited incidents and research showing agent‑driven expansion of attack surface and investigation blind spots (Palo Alto).
Advisories and active exploitation
Unit 42 reported two critical, actively exploited zero‑days in Ivanti EPMM (CVE‑2026‑1281, CVE‑2026‑1340; CVSS 9.8) stemming from unsafe input handling in legacy bash scripts triggered via Apache RewriteMap, enabling unauthenticated RCE through crafted GET requests. Observed attacks drop JSP web shells, monitoring agents, cryptominers, and reverse shells, with the potential for appliance persistence and push of malware to managed devices. Ivanti has released patches and detection scripts; CISA has added CVE‑2026‑1281 to KEV. Immediate patching, hunting for provided indicators, and appliance forensics are advised (Unit 42).
CISA published a critical advisory for multiple Honeywell CCTV devices (CVE‑2026‑1670; CVSS 9.8) exposing an unauthenticated API that can change the password‑recovery email, enabling account takeover and unauthorized feed access. CISA recommends isolating devices, minimizing Internet exposure, deploying secure remote access, and contacting Honeywell for patches (CISA).
Google alerted that a High‑severity Chrome zero‑day (CVE‑2026‑2441), a use‑after‑free in the CSS engine, is exploited in the wild. Fixes are available (Windows/Mac 145.0.7632.75/76; Linux 144.0.7559.75+), with technical details withheld pending broad patch adoption (CSO Online).
CISA also added four entries to the Known Exploited Vulnerabilities catalog—spanning a legacy ActiveX RCE, a Zimbra SSRF, an anti‑ransomware file‑upload flaw, and the Chrome CSS bug above—triggering federal remediation timelines and signaling higher operational risk beyond government networks (KEV).
Appliances, firmware, and developer ecosystems under pressure
Mandiant and Google Threat Intelligence detail UNC6201’s exploitation of a critical Dell RecoverPoint for Virtual Machines flaw (CVE‑2026‑22769), using default Tomcat Manager credentials to upload WAR files, deploy a SLAYSTYLE web shell, and run commands as root. The actor replaced BRICKSTORM with a native AOT‑compiled C# backdoor dubbed GRIMBOLT, and used persistence via modified startup scripts. Investigators also observed ESXi tradecraft such as “Ghost NICs” and Single Packet Authorization rules. Dell has published remediations; Mandiant/GTIG provide IOCs, YARA, and forensic guidance (Google TI). Why it matters: virtualization appliances often lack EDR coverage, allowing durable access if left unpatched.
Kaspersky reported Keenadu, a firmware‑level Android backdoor implanted during the build process and observed since August 2023, with deep hooks via libandroid_runtime.so that inject into every app process. Operators deployed modular payloads largely for ad fraud, with guardrails (for example, avoiding Chinese locales, requiring Play Services) and delayed activation to evade analysis. Telemetry indicates thousands of affected devices, including certain Alldocube models via compromised OTA; remediation may require reflashing clean firmware (The Hacker News).
Defenders and developers also saw supply‑chain and IDE risks addressed and exposed. Notepad++ introduced a “double‑lock” update mechanism in v8.9.2—independently validating the signed installer and a signed XML manifest—while removing libcurl.dll to reduce DLL side‑loading and tightening plugin execution policies, following a months‑long update‑infrastructure compromise discovered in December 2025 (BleepingComputer). Separately, researchers disclosed multiple high‑ to critical‑severity flaws in popular VS Code extensions (Live Server, Code Runner, Markdown Preview Enhanced, and a pre‑0.4.16 XSS in Microsoft Live Preview), enabling local file theft and potential RCE via malicious config snippets; recommended mitigations include pruning unnecessary extensions and avoiding untrusted configuration (BleepingComputer).
In OT, Dragos documented a sharp rise in ransomware targeting industrial organizations in 2025, with 119 groups and 3,300 victims, commonly entering via remote‑access portals and virtualization services and then impacting OT‑adjacent systems such as ESXi. The report stresses OT‑specific visibility, credential protections, segmentation, vendor‑tunnel governance, and tested recovery plans to reduce downtime (Infosecurity).