☕️ Polish health software breach hits 19M

Hi there, this is your daily ☕️ Cyberpresso.

In today's Cyberpresso:

🏥 Polish health software breach hits 19M

🔓 WordPress plugin flaw exposes 600K sites

🛡️ CISA: Patch exploited AI flaw now

🖥️ Exploit code hits Microsoft SCCM flaw

🔒 GitLab patches critical flaw that let attackers delete public projects

Plus: 💡 4 strategies & tactics, 🎁 5 other news you might like, 🧰 6 tools, and 📚 5 papers.

🏥 Polish health software breach hits 19M LINK

Polish authorities are investigating a cyberattack on healthcare software provider MyDr that may have exposed data belonging to nearly 19 million people and more than 12,000 medical facilities across the country.

MyDr said it identified and removed the cause of the incident but disclosed no details about the vulnerability or how attackers broke in; officials say hackers obtained unauthorized access to historical data held in MyDr systems through April 2024.

MyDr's software connects providers to P1, Poland's nationwide health platform for electronic prescriptions and referrals; as a precaution the e-Health Center is replacing digital certificates, though authorities found no evidence the certificates were stolen or misused.

🔓 WordPress plugin flaw exposes 600K sites LINK

A flaw in WordPress's Forminator form plugin lets attackers upload malicious files and hijack sites, so update to version 1.56.2 immediately.

🛡️ CISA: Patch exploited AI flaw now LINK

CISA is ordering federal agencies to patch an actively exploited code-injection flaw in Ray, the open-source AI framework maintained by Anyscale that distributes machine-learning workloads across clusters, or stop running the software by August 20.

The bug, CVE-2025-62593, lets an attacker who has never logged in make a vulnerable Ray deployment run commands of their choosing, and unusually it can be reached through an ordinary web browser like Firefox or Safari rather than direct network access.

Anyscale has fixed the issue in Ray version 2.52.0, leaving operators on earlier releases exposed; CISA lists any ransomware use as "unknown," so the immediate risk is unauthorised code execution on clusters holding proprietary models, training data, and cloud credentials.

🖥️ Exploit code hits Microsoft SCCM flaw LINK

Working proof-of-concept exploit code is now public for a critical remote code execution flaw (CVE-2026-47301) in Microsoft Configuration Manager (SCCM), letting a low-privileged domain user gain SYSTEM-level control of a Primary Site Server.

The exploit chain, published by researcher Omri Baso, uses a malicious CAB file with path-traversal sequences to write attacker DLLs into the Configuration Manager bin\X64 directory, which the elevated SMS_EXECUTIVE service then loads via DLL proxying.

Attackers must first identify the Primary Site Server, inferable from Full Control or GenericAll permissions on the System Management container; defenders should apply Microsoft's update and watch for changes to the RID 500 Administrator account and suspicious DLLs.

🔒 GitLab patches critical flaw that let attackers delete public projects LINK

GitLab has patched a critical flaw in its GraphQL interface, tracked as CVE-2026-19478, that lets unauthenticated attackers remotely modify or delete public projects and user data on Community Edition and Enterprise Edition installations.

Rated critical (CVSS 9.4), the code injection bug abuses a GraphQL directive so that, under specific conditions, a remote attacker with no account or user interaction can perform unauthorized actions against public GitLab resources over the network.

Fixed versions 19.2.4, 19.1.6, 19.0.8, and 18.11.11 shipped August 17, 2026; GitLab.com and Dedicated are already patched, and admins should upgrade self-managed instances and review audit logs for deleted projects or suspicious GraphQL activity.

💡 Strategies & Tactics

> Video Call Exploit Chains Two Flaws in Unisoc Modems: Researchers showed that chaining two Unisoc modem flaws lets an attacker seize Android kernel control after the victim answers a malicious video call.

> Google’s open-source HEIR lets AI work with data it can’t see: Google's HEIR compiler converts AI models to run on encrypted data, letting healthcare and finance apps analyze sensitive information without ever exposing its contents.

> Behavioral Malware Analysis: Investigating a Multi-Stage Malware Sample Inside an Isolated Lab: Run malware in an isolated lab and watch it live to capture the attack behaviors static analysis misses, then turn those observations into detection rules defenders can actually use.

> There’s a libcurl.dll in my system32: A stray libcurl.dll flagged by a scanner must be updated by whichever app installed it, not by the curl project itself.

Other news you might like

• Geekom admits to shipping malware-laced network drivers for AMD mini PCs, company responds with guidance, removes malicious packageLINK

• Octagon Android Bot Uses Hidden VNC and Accessibility Overlays to Steal Crypto Wallet CredentialsLINK

• C2Looper v2 Uses GitHub Repositories as Full Command-and-Control Infrastructure.LINK

• Operation ASTERIX Uses Vishing and Fake Crypto Wallet Apps to Steal Seed PhrasesLINK

• VMware Syslog Path Traversal Becomes Root RCE, Persistent SSH Access and ESXi RansomwareLINK

🧰 Trending tools

Kastra: a runtime authorization layer that enforces policies on AI agents before actions execute, blocking unauthorized tool use and data exposure across major frameworks.LINK

Perfai Security: an automated tool that scans AI-generated apps from Replit, Lovable, Cursor, and Claude Code for access control vulnerabilities, fixing them with a single prompt.LINK

Constellation Gate AI: routes AI agent traffic through a gateway that blocks prompt injections, scans for secrets, logs audit trails, and cuts token costs 20-40% via compression and caching.LINK

TailMux: lets you connect to multiple Tailscale tailnets simultaneously on macOS and Linux by running isolated embedded nodes per profile, routing by hostname without switching accounts or VMs.LINK

Lunen.ai: an AI automation tool that logs every action taken and requires approval on risky steps, giving teams usability without sacrificing oversight.LINK

qsa.sh: scans your server's public IP with naabu, nmap, and nuclei to reveal open ports, service versions, and known CVEs in about 30 seconds, no signup required.LINK

📚 Trending papers & reports

AI agent "skills" can be secretly tweaked so their normal network traffic patterns leak private user information to an eavesdropper, even though the tweaks pass existing AI security audits.LINK

Retrieval poisoning defense retrains only a tiny sliver of a search system's encoder, under 1% extra parameters, so poisoned lookalike content gets filtered out before it ever reaches the chatbot's answer, with no added per-query slowdown.LINK

DeepSeek's AI assistant can be tricked by hidden instructions buried in files or text, with attacks succeeding up to ~26% of the time, showing untrusted content needs stronger checks before it reaches sensitive actions.LINK

AI-built network attack simulations matched traditional training methods with a 94.5% success rate at finding security holes, while running about 25,000 to 50,000 times faster.LINK

Next-gen network security flags malicious traffic in live 6G-style test networks while cutting the data needed to spot attacks by 80%, and pinpoints exactly which traffic traits give hackers away.LINK

See you tomorrow for a new dose of ☕️ Cyberpresso!