This week’s security news has a familiar rhythm: a system looks fine at first, attackers quietly probe, and then a trusted component turns into the pathway. Researchers and agencies report that threats are getting faster and, in some cases, cheaper to execute—especially when AI helps automate parts of the attack.
Below is a practical recap of the most important developments, what they mean, and where teams should focus next.
Threat of the Week: AI-Powered PLC Attacks
U.S. agencies warned of AI-Powered PLC Attacks targeting internet-exposed Siemens S7 Series programmable logic controllers (PLCs). These PLCs are used across water systems, energy, manufacturing, and other critical infrastructure environments.
The core issue is not theoretical. The agencies described this as an active threat. Attackers can abuse weakly secured PLCs to disrupt industrial processes, trigger safety incidents, cause downtime or equipment damage, and potentially expose sensitive data. There’s also a risk of compliance violations and knock-on effects across interconnected systems.
How the activity was described:
- Threat actors used legitimate scanning services (such as Censys and ZoomEye) to locate Siemens S7 PLCs exposed to the internet or insufficiently segmented networks.
- After identifying targets, they deployed AI-generated exploit scripts that could be disguised as normal monitoring-related activity.
- Actors were reportedly refining exploitation techniques against specific PLC models to improve their ability to compromise devices.
- For preparation, they used read access to understand the environment before attempting later actions that could cause operational disruption.
At the time of reporting, it was not known who was behind the campaign.
GitLab vulnerability exploited quickly after disclosure
A newly disclosed GitLab security flaw moved from “announced” to “actively exploited” within days. According to watchTowr, CVE-2026-19478 (CVSS 9.4) is a code injection issue that can allow an unauthenticated attacker to modify or delete publicly accessible GitLab projects and rewrite data under certain conditions.
What makes this especially concerning is that it can happen without credentials, without user interaction, and without relying on hard-to-achieve setup requirements.
If you run GitLab instances, prioritize patching and review whether any systems were reachable from untrusted networks during the exposure window.
Trojanized npm packages deliver RedC2 4.0
Researchers also reported a cluster of trojanized packages on npm—specifically 14 malicious entries designed to look like legitimate calendar and streak utilities. Instead of merely running expected functions, these packages were built to stealthily deliver an AI-powered Linux backdoor called RedC2 4.0.
The reported RedC2 4.0 capabilities include:
- Surveillance activities
- Credential theft
- Payload loading
- Mass-operation features for scaling actions
RedC2 4.0 was marketed on cybercrime forums as a cross-platform toolkit for Windows, macOS, and Linux, and it was described as an evasion-focused command-and-control framework. The reports also mention a forum user name (“MarlboroMan”) linked to early promotion.
New card fraud technique with physically expired Visa cards
Another item from the research track: a “Zombie Card” method that can bypass cryptographic checks to complete contactless payments using physically expired Visa cards.
Researchers described using a smartphone relay setup to modify the expiration date fed to the point-of-sale terminal without breaking the card’s cryptography. In reported lab results, transactions succeeded only at some banks when the team adjusted a specific CDCVM flag.
Importantly, the reporting states that there’s currently no evidence the technique has been used in the wild.
Suspected Russian clusters abuse legitimate authentication flows
Google attributed observations to three suspected Russian cyber-espionage clusters: UNC6293, UNC7005, and UNC5976. The clusters focus on targeting individuals across academia, aerospace and defense, government entities, and think tanks in both Europe and the U.S.
The key pattern: the attackers reportedly use legitimate authentication workflows to single out victims, while also running persistent, adaptive phishing campaigns and using sophisticated social engineering.
One cluster, UNC7005, has been associated with CaptiveCrunch—an approach that targets captive Wi-Fi portals (for example, at hotels, conference centers, and airports) to redirect users to attacker-controlled infrastructure and capture credentials.
A separate report also claims that three managed service providers were likely compromised to support captive-portal hijacking through a supply-chain route.
Cloudflare Workers Spectre attack leaks JWT
Researchers described a remote Spectre-style attack against Cloudflare Workers. The reported impact is the leakage of a JSON Web Token (JWT) from a co-located Worker in a production environment.
According to the study, data leakage occurred at a low throughput rate (up to 12 bits per second), but at a frequency described as much higher than a previous 2021 demonstration.
The researchers also indicated that the technique used amplification methods combined with a remote timing server.
Cl0p targets PTC Windchill with a tailored web shell
ReliaQuest reported that the Cl0p ransomware group used a bespoke web shell after exploiting a critical security weakness in PTC Windchill and FlexPLM servers. The web shell is described as being specifically built for enterprise Product Lifecycle Management (PLM) environments.
Beyond basic access, the shell is described as capable of mapping sensitive “vault” data, decrypting credentials from the Windchill keystore, and executing further code using a custom Java class loader.
This is not presented as a one-off behavior. The reporting notes that Cl0p has previously deployed custom shells after exploiting SQL injection flaws in Accellion and MOVEit Transfer file transfer software.
As of August 12, 2026, the ransomware group began releasing alleged victims’ full names, and the report suggests that over 40 organizations have been targeted.
Unisoc firmware flaw plus another RCE issue can enable kernel access
Researchers disclosed a new unpatched issue in Unisoc T612 modem firmware. They also stated that—when combined with a previously reported remote code execution flaw (also unpatched)—the chain could allow attackers to obtain elevated access to the Android kernel on affected devices.
The described path includes delivering a malicious payload to the phone’s modem via the earlier RCE vulnerability, then placing a video call to the device that the victim would need to answer. The reporting quotes SSD Secure Disclosure about disabling protections in a specific memory region to reach unrestricted read/write capabilities, potentially enabling local privilege escalation and the ability to modify kernel code.
Trending CVEs: patch quickly and focus on what’s actively targeted
As always, the gap between patching and exploitation appears to be shrinking. This week’s “urgent-first” advice is simple: patch what you run, and treat the most widely used or currently exploited issues as immediate priorities.
The recap lists many high-severity CVEs and includes examples such as:
- CVE-2026-15748 (Forminator Forms)
- CVE-2026-15826 (User Profile Builder)
- CVE-2026-73570 (Zimbra)
- CVE-2026-32475 (Elementor Pro)
- CVE-2026-64849 (MLflow)
- CVE-2026-19478 (GitLab)
- CVE-2026-65922 (JFrog Artifactory)
- CVE-2026-25895 (FUXA)
It also highlights long sets across major products and browser ecosystems, including multiple Mozilla Firefox and Thunderbird entries, and several items for Google Chrome. The list further includes CVEs for Atlassian Bamboo Data Center, Splunk, Spring Security (via UnboundID LDAP server issues), Keycloak, and several others.
Even if you don’t patch everything this week, use the list to drive triage: identify exposed internet-facing services, check asset inventories, and prioritize systems with active exploitation signals.
Leaked credentials and platform risks
Beyond vulnerabilities, this week also brought reminder-level failures in the “credential and access” category.
Stripe API keys leaked for hundreds of merchants
A dataset published on a data-trading forum reportedly includes live Stripe API keys for 659 merchants, along with around 35 GB of customer and payment-related data pulled from those accounts. The reporting emphasizes that a Stripe secret key isn’t just a dashboard login—it can provide programmatic access, including reading customer records, creating charges, issuing refunds, and changing payout destinations.
The same report suggests that a subset of the exposed keys (listed as 519 in a row) could potentially both receive and move funds on the account record.
AWS keys exposed with full administrative rights
Truffle Security reported finding 768 leaked corporate AWS keys that carry full control. The scan reportedly verified thousands of unique key pairs across public findings, including cases where keys authenticate successfully and where some keys provide AdministratorAccess or root-level control.
The broader risk here is time and rotation: the median live leaked key reportedly had not been rotated for years.
Bottom line: attackers rarely need everything to fail
If there’s one takeaway from this week’s mix of industrial alerts, quick-to-exploit CVEs, and real-world credential exposure, it’s that attackers often don’t need a “perfect” chain. One exposed service, one trusted shortcut, or one overlooked dependency can be enough to get started—and the rest can follow.
So instead of only asking what the next big threat is, teams can ask a more useful question: what do we still assume is safe? That habit tends to surface weaknesses earlier—before automation, AI assistance, or opportunistic exploitation turns them into incidents.
Source: https://thehackernews.com/2026/08/weekly-recap-ai-powered-plc-attacks.html
