Over 100 hacked WordPress sites redirect visitors to fake Cloudflare/Google CAPTCHA pages using “ClickFix” to execute Lumma, Rhadamanthys, Epsilon Red, and XMRig payloads, researchers say.

A global campaign dubbed ShadowCaptcha is abusing more than 100 compromised WordPress sites to funnel visitors to counterfeit CAPTCHA checks that coerce “ClickFix” actions and launch malware—ranging from info-stealers and ransomware to crypto miners. Researchers link some infrastructure to Help TDS–style redirection and malicious plugins that masquerade as WooCommerce. The Hacker NewsGoDaddy

A large-scale attack wave called ShadowCaptcha is redirecting users from more than 100 hacked WordPress sites to fake Cloudflare or Google CAPTCHA pages that trigger multi-stage malware installs, including Lumma and Rhadamanthys info-stealers, Epsilon Red ransomware, and XMRig coin miners, according to new research published August 26, 2025.

• What’s new: ShadowCaptcha leverages compromised WordPress sites to run malicious JavaScript that sends visitors into a redirection chain ending on phony CAPTCHA pages. From there, victims are prompted to either paste a pre-copied command in Windows Run or save and execute an HTA file—both paths resulting in malware execution. The Hacker News
• Scale: Researchers have observed 100+ infected sites, with concentrations in Australia, Brazil, Italy, Canada, Colombia, and Israel across technology, hospitality, legal/finance, healthcare and real-estate sectors. The Hacker News
• Initial access: How the WordPress sites were breached varies; investigators have medium confidence that attackers abused vulnerable plugins and, in some cases, stolen admin credentials. The Hacker News
• Related ecosystem: The campaign overlaps with traffic distribution systems (TDS) behavior seen in Help TDS, which has used a malicious plugin named “woocommerce_inputs” to redirect traffic and harvest credentials on thousands of sites. GoDaddy

“The campaign blends social engineering, living-off-the-land binaries, and multi-stage payload delivery to gain and maintain a foothold in targeted systems.” — Researchers credited by Israel’s National Digital Agency. The Hacker News

“The compromised ClickFix page copies a malicious command to the clipboard without interaction, relying on users to paste and run it unknowingly.” — Researchers describing the technique. The Hacker News

“Help TDS has evolved into a malware-as-a-service offering, with a malicious WooCommerce-named plugin installed post-compromise via stolen credentials.” — Denis Sinegubko, GoDaddy Security. GoDaddy

“ShadowCaptcha shows how a simple CAPTCHA lure can escalate into data theft, crypto mining, or full ransomware impact—often with mshta/msiexec abuse and vulnerable drivers for stealth and speed.” — El Mostafa Ouchen, cybersecurity author and analyst.

---

Technical Analysis

Attack chain & lures. Compromised WordPress pages inject JavaScript that redirects to counterfeit Cloudflare/Google CAPTCHA portals. The pages use ClickFix instructions to:

1. open Windows Run and paste an attacker-supplied command (copied via navigator.clipboard.writeText), launching MSI or HTA payloads via msiexec.exe/mshta.exe; or
2. save the page as an HTA and execute locally. The Hacker News

Payloads. Observed families include Lumma and Rhadamanthys (stealers), Epsilon Red (ransomware), and XMRig miners (with configs sometimes fetched from Pastebin). Some runs drop a vulnerable driver (WinRing0x64.sys) to manipulate CPU registers for higher mining yield. The Hacker News

Defense evasion. Pages implement anti-debugger checks to block browser dev tools inspection and use DLL side-loading to execute under trusted processes. The Hacker News

Possible delivery infra. Research into Help TDS documents a malicious “woocommerce_inputs” plugin used by attackers (not from the legitimate WooCommerce project) to redirect traffic, filter by geography, and exfiltrate credentials—capabilities that can dovetail with ShadowCaptcha’s redirection-first model. GoDaddy

MITRE ATT&CK (indicative):

• Drive-by Compromise (T1189) via compromised sites and forced redirects.
• User Execution (T1204) through ClickFix-guided Run/HTA steps.
• Signed Binary Proxy Execution (T1218) using mshta.exe / msiexec.exe.
• Hijack Execution Flow: DLL Side-Loading (T1574.002).
• Valid Accounts (T1078) for stolen WordPress admin credentials.
• Exploitation for Privilege Escalation (T1068) via vulnerable driver abuse.

---

Impact & Response

Who’s at risk:

• Site visitors—credential theft, data exfiltration, ransomware execution, resource hijacking for mining.
• Site owners—reputation damage, SEO penalties, blacklisting, potential legal exposure for unsafe platforms. The Hacker News

Immediate actions:

• Users: do not paste/run commands from web pages; block HTA where feasible; run EDR; scan for Lumma/Rhadamanthys/Epsilon Red; check for unauthorized drivers.
• Admins: audit WordPress for unknown plugins (e.g., faux WooCommerce names), remove malicious injections, rotate credentials, enforce MFA, and patch core/plugins; review outbound redirects and logs; WAF/EDR rules for mshta/msiexec misuse; disable Pastebin-fetched configs at egress. The Hacker NewsGoDaddy

Potential regulatory angle: Sites handling personal data may face privacy/consumer-protection scrutiny if inadequate security controls facilitated malware delivery to visitors.

---

Background

The disclosure follows GoDaddy’s deep-dive on Help TDS, active since 2017, which arms affiliates with PHP templates and a malicious plugin to monetize hijacked traffic (tech-support scams, dating/crypto/sweepstakes), including fake CAPTCHA gates to evade automated scans. ShadowCaptcha adopts similar redirection motifs while expanding to stealers/ransomware/miners. GoDaddy

---

What’s Next

Researchers are continuing to track infrastructure and plugin variants, while urging WordPress operators to harden authentication, prune legacy/vulnerable plugins, and monitor for ClickFix-style clipboard abuse. Expect IOCs and cleanup guidance to roll out via security vendors and national agencies as investigations continue. The Hacker News

Governor’s Technology Office says no evidence of PII exposure as FBI joins probe; DMV, agency websites and phone lines saw disruptions while emergency 911 remained available.

RENO, Nev. — Aug. 25, 2025. Nevada paused in-person services across state agencies on Monday while technicians worked to recover from a “network security incident” first detected early Sunday, according to a memo from the Governor’s Office. Officials said certain websites and phone lines were intermittently unavailable during recovery, but emergency call-taking remained online. Carson Now

What happened

• The state identified a network security incident early Sunday, Aug. 24 and initiated 24/7 recovery and validation of systems. Agencies were instructed to suspend in-person counters Monday “to minimize disruption” while restoration proceeded. Carson Now
• As of Monday afternoon, NV.gov and several agency sites (including the Department of Public Safety) experienced outages or slow response. StateScoopCarson Now
• The FBI’s Las Vegas field office confirmed it is assisting the state’s investigation. KSNVLas Vegas Review-Journal
• No evidence of compromised personally identifiable information (PII) has been found so far, the state said. Carson NowKRNV

Services and public impact

• DMV: Offices were closed Monday; appointments set for that day will be honored as walk-ins over the next two weeks, the agency said. KRNVKSNV
• Public safety lines: Nevada Highway Patrol / State Police dispatch phone lines were down Sunday but were back in service Monday; 911 remained available statewide. Carson NowLas Vegas Review-Journal
• Other agencies: Notices from departments, including Nevada State Parks, cited website disruptions and operational adjustments (e.g., cash-only fees at park entrances). Nevada State Parks

What officials are doing

The Governor’s Technology Office said it is working “continuously with state, local, tribal, and federal partners,” using temporary routing and operational workarounds to maintain public access where feasible, and validating systems before returning them to normal. Residents were warned to be wary of phishing or payment scams purporting to be from state agencies. Carson Now

---

Technical context (what this likely means, without speculating beyond facts)

Scope and blast radius. The pattern of symptoms—intermittent web/phone outages across multiple agencies, temporary DMV service suspension, and selective restoration of dispatch lines—suggests disruption at shared network and application tiers rather than isolated end systems. That could include identity infrastructure, routing/firewall layers, telephony integration (SIP trunks, call routing), and public-facing web gateways. (This is analysis; officials have not yet attributed a cause.) StateScoopCarson Now

What we don’t know yet. Nevada has not announced whether this was a ransomware intrusion, DDoS, supplier compromise, or an internal misconfiguration. Investigators typically review authentication logs, privileged access activity, VPN/SSO telemetry, firewall and proxy events, and VoIP call-control logs to determine initial access, lateral movement, and exfiltration. The FBI’s involvement indicates the state is preserving evidence and coordinating across jurisdictions. KSNVLas Vegas Review-Journal

Why services come back in phases. Agencies are restoring in waves because modern state IT relies on centralized identity (SSO), shared network segments, and common telecom/web platforms. Best practice is to isolate, rebuild, and validate each dependency (DNS, IdP, MDM/EDR, call routing, web apps) before reopening public counters—hence Monday’s pause in person services. Carson Now

---

Guidance for residents (from the state + security best practice)

• Use official channels and be patient with intermittent outages; retry later if a site or line times out. Carson Now
• Treat unexpected calls/emails requesting payment or credentials as suspect; the state won’t ask for your password or bank details by phone or email. Carson Now
• For emergencies, call 911; for roadside assistance, use restored NHP dispatch lines or 911 if needed. Carson Now

---

What to watch next

• Attribution & root cause: whether investigators tie the outage to criminal intrusions (e.g., ransomware or help-desk social engineering seen in other states) or to non-malicious failures. StateScoop+1
• Data-exposure update: officials currently report no PII evidence; that assessment could change after forensics (if indicators of exfiltration emerge). Carson Now
• Restoration cadence: agency-by-agency reopenings and website recoveries, including NV.gov and DPS properties. StateScoopCarson Now

---

Sources

Governor’s Office memo via Carson Now; updates on services and scam warnings. Carson Now
KRNV/News4 Reno; DMV accommodations, outage timeline. KRNV
KSNV/News3 Las Vegas; FBI assistance confirmation, service pause. KSNV
StateScoop; NV.gov and DPS site impact; weekend-to-Monday outage context. StateScoop
Carson Now update; dispatch lines status Sunday vs. Monday. Carson Now
Nevada State Parks; closure notice and cash-only operations. Nevada State Parks

AI isn’t “one giant model for everything” anymore. The fastest progress is coming from specialized models—each excellent at a narrow slice of the problem—and smart integration that routes the right task to the right model. Here’s a clear, practical tour of the landscape and how to use it.

Why specialization beats one-size-fits-all

• Latency & cost: Smaller or task-specific models respond in milliseconds and are cheap to run; giant generalists aren’t.
• Accuracy on niche tasks: A focused vision model or segmenter will often beat a general LLM + prompt tricks.
• Deployment reality: Some workloads must run on device (privacy, offline) or at the edge (robots, cameras).
• Composable systems: Orchestrating multiple models lets you blend strengths—reason with one, perceive with another, act with an agent.

The roster (what each acronym really means)

🔹 LLM — Large Language Model

• What it is: A generalist text model for reasoning, content generation, coding help, retrieval-augmented Q&A.
• Strengths: Broad world knowledge, chain-of-thought reasoning, tool use via function calling.
• Limits: Slower and costlier than small models; can hallucinate; not great at fine visual detail.

🔹 LCM — Latent/Lightweight Consistency Model (compact diffusion)

• What it is: A diffusion-style model reworked for very fast image generation/upscaling.
• Strengths: Few inference steps → near-real-time visuals; great for product mockups, ads, thumbnails.
• Limits: Narrow domain (images/video); text/logic still needs an LLM.

🔹 LAM — Language/Logic Agents

• What it is: Planners/executors that call tools, browse, write code, schedule jobs, and evaluate results.
• Strengths: Turns model outputs into actions; automates multi-step workflows with guardrails.
• Limits: Needs good tools, memory, and evaluation loops; careless agents can “run away.”

🔹 MoE — Mixture of Experts

• What it is: A big model built from many “experts”; a router activates only a few per token.
• Strengths: Scales capacity without paying the full compute cost every step; good for multilingual/heterogeneous tasks.
• Limits: Harder to train/serve; quality depends on good routing.

🔹 VLM — Vision-Language Model

• What it is: Multimodal models that read images (and often video) + text.
• Strengths: Screenshot Q&A, chart understanding, document analysis, UI testing, visual troubleshooting.
• Limits: Still learning fine text in images, small fonts, edge cases; may need OCR aids.

🔹 SLM — Small Language Model

• What it is: Compact LLM (10B parameters or less) for edge/on-device work.
• Strengths: Low latency, private by default, runs on laptops/phones/IoT; great for autocomplete and local assistants.
• Limits: Narrower knowledge, weaker long-form reasoning; often paired with retrieval.

🔹 MLM — Masked Language Model (e.g., BERT-style)

• What it is: Pretraining objective that predicts missing tokens; great encoder for classification/search.
• Strengths: Semantic search, topic labeling, PII detection, entity extraction; fast and stable.
• Limits: Not generative; pair with an LLM when you need prose or code.

🔹 SAM — Segment Anything Model

• What it is: Foundation segmentation for images; pick out objects, regions, people—no labels needed.
• Strengths: Annotation at scale, medical pre-segmentation, retail shelf parsing, industrial inspection.
• Limits: Doesn’t “understand” the object class; combine with a classifier/VLM for semantics.

---

Quick chooser: which model for which job?

Goal	Best fit	Why
Long answers, reasoning, coding help	LLM	Broad knowledge + tool use
Instant images or edits	LCM	Few steps → fast + cheap
Automate multi-step tasks	LAM (agent)	Plans, calls APIs, checks results
Scale quality across domains	MoE	Capacity without full compute
Screenshot / PDF / chart Q&A	VLM	Multimodal grounding
Private, on-device assistant	SLM	Low latency + privacy
Search, classify, extract entities	MLM	Strong encoder semantics
Cut objects out of images	SAM	Robust, label-free segmentation

How they work together (a simple blueprint)

User request → Router → Orchestrator (Agent) → Tools/Models → Verifier → Answer

1. Router tags the task (vision, search, segmentation, write).
2. Agent (LAM) plans steps and calls:
   • VLM to read a screenshot,
   • SAM to isolate a component,
   • MLM to extract fields,
   • LLM/SLM to explain or draft,
   • LCM to render a visual.
3. Verifier/critic (could be a second small model or rules) checks safety, facts, or formatting.
4. Response is returned; artifacts (images, JSON) are attached.

This “specialization + integration” pattern beats any single model on speed, cost, and reliability.

Design trade-offs you’ll actually feel

• Latency: SLMs and LCMs are sub-second; large LLMs are not.
• Privacy: On-device SLM + local VLM can keep data off the cloud.
• Accuracy: Domain tasks (vision, segmentation, extraction) usually win with VLM/SAM/MLM over prompting a general LLM.
• Cost control: Use SLM/MLM for 80% of routine work; escalate to a larger LLM only when needed.
• Maintenance: More moving parts → add observability (per-model metrics, routing logs, error budgets).

---

Evaluation playbook (keep it simple)

• Define slices: e.g., “OCR-heavy PDFs,” “charts,” “legal text,” “UI screenshots.”
• Pick metrics: EM/F1 for extraction (MLM), IoU for segmentation (SAM), latency & cost per call, human preference for LLM outputs.
• A/B the router: Measure when it sends tasks to the “expensive” model—can a small model handle it?
• Guardrails: Safety filters, citation checks (for RAG), and a lightweight self-check pass on critical outputs.

---

Three mini-patterns you can borrow

1. Help desk with eyes: VLM reads user screenshots → SAM crops error dialog → LLM writes the fix; average handle time drops, fewer back-and-forths.
2. Catalog cleanup: SAM segments product photos → VLM describes → MLM tags → LCM generates clean hero images.
3. Private coding copilot: SLM runs on the developer’s laptop (context from local repo) → MoE backend only for hard refactors.

Getting started (no drama, just steps)

1. Map your top 5 tasks and their constraints (privacy, latency, budget).
2. Start with one specialist beside your LLM (e.g., VLM for screenshots, or MLM for extraction).
3. Add a tiny router (heuristics at first) and log decisions.
4. Introduce an agent once you have 2–3 tools to chain.
5. Instrument everything: latency, cost, success rate, fallback counts.
6. Iterate—promote frequent fallbacks to first-class tools, demote what you don’t use.

---

The bottom line

The future of AI isn’t a bigger hammer. It’s a toolbox:

• LLMs for reasoning,
• VLM/SAM for seeing,
• MLM for knowing,
• LCM for drawing,
• SLM for speed and privacy,
• LAM to coordinate,
• MoE to scale.

Specialization + integration is how you get real-world performance.

“The future of AI isn’t a bigger model—it’s a better orchestra.
LLMs reason, VLMs see, SAM segments, MLMs extract, SLMs protect privacy at the edge, and agents coordinate the flow.
Real performance comes from routing the right task to the right specialist and measuring the system end-to-end.”
— El Mostafa Ouchen, cybersecurity author and analyst