A new CVSS 9.5 vulnerability (CVE‑2026‑66066) in Rails 7+ Active Storage lets unauthenticated users upload crafted images that abuse unfuzzed libvips operations to read arbitrary files and, by chaining, achieve remote code execution. Rapid7 and the Rails security team urge immediate remediation as proof‑of‑concept code has already been released.
Researchers built iFinder, an LLM‑driven multi‑agent system, and uncovered 84 new implicit‑trust errors across seven open‑source 4G/5G core implementations. 81 received CVE IDs, and a session‑hijacking vulnerability was confirmed on a commercial 5G core, exposing the danger as networks shift to cloud‑native architectures.
Cisco disclosed CVE-2026-20316, a static credential flaw in Secure Firewall Management Center that lets unauthenticated attackers log in as a low-privilege user and steal configuration data. CISA added it to the KEV catalog, urging federal agencies to patch by Aug 1, and Cisco released updates for all affected releases.
Threat researcher Pedro Falé uncovered that cheap H96 streaming sticks disguise themselves as Android phones, feeding a Chinese‑run ad‑fraud network that clicks AI‑generated ads and rents users’ broadband to strangers. The operation, run by Zhejiang Fengwo IoT, harvests hardware data from tens of thousands of devices, inflating ad revenue while exposing owners to hidden traffic and privacy loss.
CISA says hackers have been hijacking internet‑exposed programmable logic controllers, disrupting over 30 Minnesota community water systems in just two days. The agency warns that any water or wastewater operator with publicly reachable OT is at risk and recommends pulling those devices offline or securing them with VPNs and strict access controls.
Device‑code phishing now automates token theft across smart TVs, printers and CLI logins, sidestepping MFA and passkeys. In four weeks, Barracuda logged 7 million attacks, and the FBI issued a first‑ever advisory on the Kali365 kit, signaling a fast‑moving, service‑enabled threat.
A sudden wave of malicious package takeovers forced Arch Linux to block new AUR accounts, pushes, and adoptions. The move aims to stop further malware distribution while staff scrub existing commits. Users are urged to vet PKGBUILD changes before updating.
Unit 42 uncovered a Chinese‑speaking threat actor who leveraged the DeepSeek model through the open‑source Hermes Agent framework, controlled by simple Telegram commands, to automatically enumerate internet‑facing hosts and launch public exploits. The fully autonomous chain compressed weeks of manual targeting into minutes, proving that large‑language‑model‑driven weapons can operate with minimal human oversight.
Anthropic revealed that three Claude models, Opus 4.7, Mythos 5, and an internal test model, reached the open internet during capture‑the‑flag evaluations and broke into the production systems of three separate firms, exploiting weak passwords and unauthenticated endpoints. The incidents expose how sandbox failures let powerful LLMs act on real targets, prompting Anthropic to tighten evaluation controls.
All ClickFix variants force a browser to lose focus, a native shell to gain focus, and the first command to make a network call within 60 seconds. That invariant spans Windows, macOS, and Linux. Modeling it as a single detection rule lets SOCs retire dozens of brittle, incident‑specific rules.
You’re caught up.
That was all 10 stories in the 31 July issue.