Critical Exploits, Vendor Controls, and AI Service Stability

Enterprise defenders faced an accelerated week: critical exploits hit business platforms, major vendors rolled out access-hardening features, and multiple organizations disclosed data losses. The Apple and macOS ecosystem saw both active exploitation and targeted spyware alerts, while the AI sector balanced reliability concerns with new provenance controls. Below are the key developments and what they mean for operations and risk.

Vendors Tighten Access and Database Security

Cloudflare’s latest update to Access makes Worker-hosted applications private by default, whether scoped to individual Workers or applied account-wide, reducing accidental exposure during rapid prototyping and deployment. According to the Cloudflare blog, Access now intercepts requests on all entry paths (custom domains, workers.dev, routes, preview URLs) before code runs, supports granular identity policies and service tokens, and propagates authenticated identity to the Worker via ctx.access for authorization and logging. The change extends to Workers for Platforms by protecting the dispatch Worker, and was enabled by Cloudflare’s FL2 proxy architecture, which separates routing from execution to enforce Access per-Worker.

Oracle introduced Database Security Central, a consolidated security console for Oracle database environments that is available at no charge until the end of February 2027. As reported by CSO Online, the tool centralizes visibility into risk, detects configuration drift, highlights privileged-access weaknesses, tracks sensitive data use, collects audit evidence, and unifies policy management. The company positions it to help teams prioritize remediation and maintain consistent enforcement amid increased threat activity and a faster monthly patch cadence.

Critical Exploitation Alerts Across Enterprise and Apple Ecosystems

Defused threat intelligence observed active targeting of CVE-2026-58231, a maximum-severity remote code execution flaw in SAP Commerce Cloud’s Data Hub Adapter extension. Per BleepingComputer, unauthenticated attackers can trigger vulnerable functions due to improper authorization and a default authentication client, enabling arbitrary code execution. Although SAP’s advisory did not mark the issue as actively exploited, honeypot hits and broad internet exposure suggest urgent patching, endpoint review, and post-exploitation monitoring are warranted for Commerce Cloud deployments.

The Netherlands’ NCSC reported active exploitation of a macOS Screen Sharing authentication bypass (CVE-2026-65400) that allows remote access to systems listening on TCP 5900; victims have seen root access followed by deployment of a Monero miner. Apple addressed the issue in recent macOS updates, and organizations are advised to patch or disable Screen Sharing if immediate updates are not feasible, as detailed by BleepingComputer. Separately, Apple sent new high-confidence threat notifications to users likely targeted by mercenary spyware; recipients are urged to verify the alert via Apple’s account portal and consider enabling Lockdown Mode, according to BleepingComputer.

Data Theft and Exposure Affect Multiple Sectors

RingCentral disclosed a July breach stemming from a “sophisticated social engineering” campaign. Analysis by Have I Been Pwned of leaked material confirmed data for about 1.6 million accounts, including names and contact details, while core platform operations were reportedly unaffected. The extortion group ShinyHunters claimed responsibility; RingCentral said the incident impacted a limited set of customers and that remediation measures are in place, per BleepingComputer. In the energy sector, Shell is investigating after the Clop group claimed to have stolen 89GB of data. The campaign is linked to exploitation of a critical input validation flaw in PTC Windchill and FlexPLM (CVE-2026-12569); advisories urged emergency patching and mitigations for affected PLM systems, according to BleepingComputer.

Researchers at Fortra Intelligence and Research Experts (FIRE) tied recent public data dumps to the ExfilSquad extortion group, confirming access to sensitive records across at least 13 organizations in government, education, financial services, and manufacturing. FIRE attributes the leaks to unauthorized read access to Microsoft D365 environments via misconfigured Microsoft Power Pages portals that incorrectly grant public read permissions through the Anonymous Users role. The data sets matched Microsoft Dataverse exports and totaled roughly 27 million records, underscoring the systemic risk of exposed portals and the need to follow table-permission guidance, per Infosecurity Magazine.

AI Reliability, Watermarking, and Agent Controls

Anthropic experienced a significant outage beginning around 21:58 UTC on August 16 that disrupted sign-ins and caused errors across multiple Claude services; Claude.ai, Claude Code, and Claude Cowork were listed as under a major outage, while Claude Console and the API remained operational. The company is investigating authentication and performance issues and posting updates to its status page, as reported by BleepingComputer.

On macOS, Jamf researchers documented a new infostealer dubbed AmnesiaStealer that advances from mere data theft to interactive session abuse. The malware’s loader harvests credentials, keychain items, and messaging data and then retrieves a stream_module to clone Chromium browser profiles and operate a hidden, headless browser on the victim’s device via the Chrome DevTools Protocol. This enables attackers to control authenticated sessions, manage tabs, and import or export cookies in real time, according to BleepingComputer.

Anthropic also detailed plans for invisible watermarking of Claude-generated text, implemented by subtly altering the model’s internal randomness to embed a statistical signature without changing the final output. Detection requires a secret key and is probabilistic, with higher reliability on longer, higher-entropy text; short samples and heavy rewrites reduce confidence. A detection API is in development to estimate whether Claude contributed to content, per BleepingComputer.

As organizations adopt AI agents using the Model Context Protocol (MCP), Cloudflare introduced Cloudflare One capabilities to detect and govern MCP traffic. Signals such as MCP-Protocol-Version and Mcp-Method allow classification of MCP over inspected HTTPS; detected requests surface in Gateway HTTP logs and can be filtered using a dedicated selector, enabling teams to allow, block, or DLP-scan JSON-RPC traffic and prevent bypass of approved MCP Server Portals. The goal is to curb shadow MCP endpoints and data exfiltration risks from nondeterministic agents, according to the Cloudflare blog.