In a detailed post on the Google blog, the Chrome team outlined a layered security architecture for agentic browsing, centered on an isolated User Alignment Critic, Origin Sets, deterministic filters, and explicit user confirmations. At re:Invent, new AI‑ and automation‑enhanced security capabilities were detailed on the AWS blog, including agentic investigations, expanded GuardDuty coverage, and identity innovations designed to speed policy generation and contain lateral movement. Together, these moves frame a day focused on proactive controls while critical advisories and confirmed intrusions reinforce the need for rapid patching and disciplined identity governance.
Agentic and cloud defenses advance
Chrome’s agentic browsing protections pair model‑level guardrails with browser controls to reduce the impact of indirect prompt injection and autonomy gone wrong. The architecture separates decision‑making between a high‑trust User Alignment Critic model that vets actions and a planner bound by Agent Origin Sets that gate read/write access to sites and tools. Deterministic filters block risky destinations and sensitive local data access, while user confirmations stand in front of sign‑ins, purchases, and other consequential steps. Operational defenses—an on‑device prompt‑injection classifier, Safe Browsing, automated red‑teaming, telemetry, and auto‑update pipelines—are designed to catch regressions and harden over time. On the cloud side, the re:Invent announcements emphasize moving detection and response from manual triage to agent‑assisted workflows, broadening signal correlation across VMs, containers, serverless, and backups and tightening identity boundaries with agent‑centric controls.
CrowdStrike expanded its identity security stack with Falcon Shield updates that bring centralized visibility and governance to autonomous AI agents alongside a native stream of first‑party SaaS telemetry into Falcon Next‑Gen SIEM. The platform discovers agents, maps delegated access and privileges back to accountable identities, flags risky configurations such as internet exposure and over‑permissioning, and correlates SaaS events with endpoint, cloud, and network telemetry in a single investigative timeline. Automated actions via Falcon Fusion SOAR—alerting owners, suspending risky agents, or disabling accounts—extend familiar playbooks for human identities to non‑human and agentic identities, addressing an attack surface increasingly dominated by malware‑free techniques like credential theft, session hijacking, and SaaS abuse.
Developer tools, governance, and automation
IAM Policy Autopilot debuted as an open‑source static analysis tool that scans Python, Go, and TypeScript code to propose functional baseline IAM policies, mapping SDK calls to required permissions and common cross‑service dependencies. It runs as a CLI or MCP server so coding assistants can request accurate policy suggestions during development and tests, helping developers iterate quickly while planning for later least‑privilege refinement. On the platform side, Application Design Center reached general availability as an AI‑enhanced, application‑centric design and lifecycle tool. It transforms natural‑language intents into deployable, multi‑product Terraform templates, tracks immutable template revisions for audit, detects drift, and integrates with App Hub and Cloud Hub to connect design with runtime operations. The approach supports GitOps and CI/CD so infrastructure changes remain versioned, reviewable, and aligned with organizational standards.
Automation also moved into research workflows as Quick Suite integrated Quick Research with Quick Flows. Teams can now schedule or trigger full research reports as steps in multi‑stage flows, generating source‑traced insights that can kick off downstream actions—updating a CRM, opening a compliance ticket, or creating a legal task. The capability aims to standardize repeatable analysis while preserving auditability and tailoring outputs through controlled user inputs.
Advisories and active exploitation
CSO Online reports that an XML External Entity vulnerability first disclosed as CVE‑2025‑54988 in Apache Tika’s PDF module has been subsumed into a broader CVE‑2025‑66516 affecting core components, with a maximum severity rating of 10.0. The flaw can trigger sensitive data exposure or unintended internal and external requests during document processing. Maintainers recommend upgrading to tika‑core 3.2.2, the standalone PDF module 3.2.2, or tika‑parsers 2.0.0 for legacy users, and caution that Tika can be a transitive dependency. Where patching is delayed, disabling XML parsing via configuration can mitigate risk; organizations that applied the earlier module‑specific fix are advised to re‑audit for the broader issue. In parallel, Infosecurity details rapid, in‑the‑wild exploitation of React2Shell (CVE‑2025‑55182), a pre‑authentication RCE in React Server Components with a CVSS 10. Observers report widespread scanning and weaponization by multiple actors, with public proof‑of‑concepts of varying quality adding noise that can mask real compromise. Providers highlight the operational risks of emergency mitigations and urge careful validation to avoid service disruption while closing exposure.
CISA added two issues to the Known Exploited Vulnerabilities Catalog: CVE‑2022‑37055 in D‑Link routers (buffer overflow) and CVE‑2025‑66644 in Array Networks ArrayOS AG (command injection). Under BOD 22‑01, federal agencies must remediate by deadlines, and CISA urges all organizations to prioritize inventory, patching, compensating controls, and enhanced monitoring. Separately, Hacker News describes active exploitation of a critical RCE in the Sneeit Framework WordPress plugin (CVE‑2025‑6389), fixed in version 8.4, where attackers create admin users and drop web shells via crafted requests to admin‑ajax.php. Guidance includes immediate updates, log audits for unexpected accounts and uploaded PHP files, and targeted blocking or monitoring of observed source IPs.
Confirmed breaches and mobile surveillance
Infosecurity reports that Marquis Software Solutions disclosed a breach affecting over 780,000 individuals after attackers exploited a SonicWall firewall vulnerability in mid‑August. Review findings indicate unauthorized access and copying of files with personal and financial data across at least 74 banks and credit unions. The company has offered credit monitoring and cited post‑incident measures, including patching firewall devices, rotating and pruning accounts, enforcing MFA on firewall and VPN access, increasing logging retention, and applying geo‑IP and botnet filtering. The investigation continues, and at disclosure time no online posting of the stolen data had been found.
Infosecurity also covers Barts Health NHS Trust seeking a High Court injunction to prevent dissemination of files stolen from an Oracle E‑Business Suite database, after a criminal group posted compressed data on the dark web. The trust says core clinical systems were unaffected, but the breached files include invoices and records with personal and supplier details. The incident fits a wider campaign tied to Oracle E‑Business Suite vulnerabilities disclosed in July, with roughly 100 organizations believed to be impacted.
On mobile, Infosecurity highlights an upgraded ClayRat Android spyware variant that expands surveillance, persistence, and remote‑control capabilities through Accessibility Service abuse, SMS privileges, keylogging, screen recording, overlays, and scripted gestures that hinder removal. Distributed across hundreds of repackaged APKs via phishing domains and file‑hosting platforms, the campaign raises risks for BYOD environments where a single infected device can facilitate data theft and unauthorized access. Recommended defenses include device‑level mobile security, strict permission and accessibility controls, MDM enforcement, and user awareness to reduce sideloading and phishing success.