Incidents and intrusions
Google’s threat team reports a coordinated data‑theft wave abusing compromised OAuth tokens tied to the Salesloft Drift connected app to access customer Salesforce instances. In its analysis, Google Cloud attributes the campaign to UNC6395 between August 8–18, 2025, detailing SOQL‑driven enumeration of Accounts, Cases, Users and Opportunities, searches for exposed secrets, and deletion of query jobs to reduce traces. Salesloft and Salesforce revoked all active tokens for Drift on August 20; Event Monitoring logs remain useful for investigation. The post includes indicators and recommends revoking and rotating exposed keys, resetting passwords, tightening connected‑app scopes, and reviewing logs for the provided IPs and user‑agent strings.
Check Point tracks a long‑con phishing operation, ZipLine, that approaches targets via public contact forms, builds weeks‑long professional exchanges under NDAs, and then delivers booby‑trapped ZIP archives. According to Check Point, payloads deploy an in‑memory backdoor, MixShell, which prefers DNS tunneling for command‑and‑control and falls back to HTTP(S). The campaign primarily targets U.S. manufacturing and supply‑chain–critical firms, with operational risks ranging from IP theft to broader supply‑chain compromise; recommended defenses include stronger scrutiny of inbound form inquiries, archive handling, and monitoring for DNS/HTTP tunneling.
A separate web‑redirect scheme compromises WordPress sites and steers visitors to convincing fake Cloudflare or Google CAPTCHA pages that instruct “ClickFix” actions. As documented by The Hacker News, the ShadowCaptcha campaign abuses clipboard and built‑in Windows tools (msiexec.exe, mshta.exe) to deliver information stealers such as Lumma and Rhadamanthys, XMRig‑based miners, and in some paths Epsilon Red ransomware. The operators rely on obfuscated scripts, anti‑debugging, and LOLBins to persist and evade detection.
Maryland’s transit agency confirmed unauthorized access affecting certain systems supporting state transport. While core services continued operating, Infosecurity reports that Mobility Paratransit temporarily could not accept new bookings or reschedule requests, real‑time updates and call center support were degraded, and some elevator services were impacted due to unavailable emergency phones. The MTA is investigating with state IT, third‑party responders, and law enforcement; the Statewide Emergency Operations Center is coordinating the response.
Unit 42 describes a wave of data‑theft extortion against luxury retail and adjacent industries that leans on voice‑based phishing and impersonation of IT support to obtain credentials or trick users into running a modified Salesforce Data Loader. Unit 42 attributes initial access and internal reconnaissance to UNC6040 and links extortion to Bling Libra (aka ShinyHunters). The activity focuses on collecting customer PII across Salesforce, SharePoint, and Microsoft 365, often leaving few forensic traces.
On mobile threats, researchers at Zimperium observed a HOOK Android banking‑trojan variant combining spyware and ransomware tactics. As summarized by The Hacker News, the malware now supports 107 remote commands, adds full‑screen payment overlays, real‑time screen streaming, fake NFC prompts, and deceptive unlock screens, and abuses Accessibility Services. Distribution includes phishing sites and bogus GitHub repositories hosting malicious APKs.
Patches and advisories
U.S. authorities added CVE‑2025‑7775, a memory‑overflow flaw affecting Citrix NetScaler, to the federal Known Exploited Vulnerabilities list based on evidence of active exploitation. CISA KEV urges prompt remediation; under BOD 22‑01, federal agencies must address KEV entries by the designated due dates, and CISA encourages all organizations to prioritize remediation.
CISA published details on multiple memory‑corruption issues in INVT engineering tools. The advisory CISA ICS lists CVE‑2025‑7223 through CVE‑2025‑7231 affecting VT‑Designer v2.1.13 and HMITool v7.1.011 in the parsing of PM3 and VPM files. Exploitation requires user interaction; CISA notes worldwide deployment, impacted critical sectors, and recommends isolating control networks, minimizing exposure, and contacting the vendor.
An improper input‑validation issue in Schneider Electric Modicon M340 controllers and communication modules can be triggered via a crafted FTP command to cause a denial of service. The republished advisory CISA ICS tracks this as CVE‑2025‑6625 and notes fixes for BMXNOE0100 (v3.60) and BMXNOE0110 (v6.80), with mitigations including disabling FTP if unused and segmenting access to port 21.
CISA also added three vulnerabilities to KEV after observed exploitation: two Citrix Session Recording issues (CVE‑2024‑8068, CVE‑2024‑8069) enabling privilege escalation or limited RCE by authenticated domain users, and a Git configuration‑parsing flaw (CVE‑2025‑48384) that can enable unintended code execution when cloning with crafted CR characters, symlinks, and hooks. See The Hacker News for the summary and remediation timeline.
Platforms and cloud
AWS introduced native remote access to IPv6 workloads: AWS Client VPN endpoints can now be configured IPv6‑only or dual‑stack across regions where the service is offered (except Middle East/Bahrain), with no separate charge for IPv6. In database resilience, Aurora DSQL added controlled chaos‑testing via the AWS FIS integration to simulate connection disruptions and regional outages, export experiment outputs to S3, and embed repeatable fault scenarios into CI/CD.
In services and response operations, Google Cloud was named a Leader in IDC MarketScape’s Worldwide Incident Response 2025 assessment, highlighting Mandiant’s integrated forensic, intelligence, and crisis‑communications model and tooling across multicloud and on‑prem environments.
Research and policy
New scrutiny of critical communications security surfaced after researchers extracted and reversed a Sepura radio’s end‑to‑end encryption and found that a 128‑bit key is compressed to an effective 56‑bit key before encryption. The analysis, covered by Schneier, argues the design choice amounts to a built‑in backdoor that undermines confidentiality for police, military, and emergency services, and calls for transparent, openly reviewed cryptography.
On cellular security, researchers from SUTD released Sni5Gect, an open‑source toolkit demonstrating practical over‑the‑air sniff‑and‑inject attacks against 5G UEs during the brief pre‑NAS‑security window. As summarized by The Hacker News, the method can crash modems, fingerprint device stacks, and downgrade connections to 4G, reviving known risks. Reported experiments achieved high injection success at short range, and GSMA acknowledged the multi‑stage downgrade as CVD‑2024‑0096.
Enterprises accelerating autonomous‑agent deployments are being urged to treat agents as a new workload class with dedicated identity, access, data, and runtime controls. In a governance blueprint, Microsoft proposes an agent registry for inventory and ownership, a seven‑layer defense model across design and runtime, and Entra Agent ID for least‑privilege, auditable access.
Operational guardrails for Model Context Protocol are also emerging. Cloudflare opened beta access to MCP Server Portals, a central gateway enforcing identity, MFA, device posture, geographic policies, and detailed logging on MCP connections, aiming to reduce risks from prompt/tool injection and confused‑deputy abuses while simplifying client configuration.
In procurement, CISA released an interactive Supplier Response Web Tool that operationalizes its Software Acquisition Guide. The tool steers buyers and suppliers through context‑relevant questions, supports secure‑by‑design principles, and produces exportable summaries to streamline due diligence and cross‑team coordination.
Finally, researchers at ESET documented PromptLock, a proof‑of‑concept ransomware that uses a local LLM (gpt‑oss‑20b via Ollama) to generate malicious Lua scripts for reconnaissance, exfiltration, and encryption across Windows and Linux samples. ESET notes the samples appear work‑in‑progress but argues the design illustrates how readily available AI models can automate and adapt extortion workflows.