Daily Cybersecurity News – August 3, 2026

What happened

Attackers drained an estimated $89 million in Bitcoin from Coldcard hardware wallets in multiple waves starting July 30, 2026, by exploiting a legacy firmware flaw that produced weak wallet seeds.

The bug, present since March 2021, caused some wallets to fall back to a deterministic software RNG instead of the hardware RNG when generating seeds. This allowed offline reproduction of private keys for addresses whose seeds lacked sufficient entropy (e.g., fewer than 50 dice rolls) or a strong BIP-39 passphrase. Galaxy Research tracked the initial automated sweep of 1,082.65 BTC ($70m) from 1,196 addresses in 41 minutes, with totals later reaching ~1,367 BTC across 4,385 addresses. Coinkite has released fixed firmware.

Who is affected

Owners of Coinkite Coldcard Bitcoin-only hardware wallets (Mk2, Mk3, Mk4, Q, Mk5 models) whose seeds were generated on affected firmware without adequate additional entropy or a strong unique passphrase.

Specifically at risk: Mk2/Mk3 versions 4.0.1 through 4.1.9, and Mk4/Q/Mk5 before the latest fixed releases (which provided only ~72 bits of entropy instead of 128). Thousands of addresses were hit; the attack remains ongoing for unmoved funds.

Why it matters

Hardware wallets are trusted as the gold standard for cold storage of high-value crypto. A multi-year entropy failure turning seeds into guessable keys undermines that trust and enables silent, automated mass theft without physical access or malware on the device itself.

Operators and holders of significant BTC balances face immediate total loss risk; the speed and scale show how quickly weak crypto primitives can be monetized once discovered.

How it could have been prevented

Immediately move any funds from potentially affected Coldcard single-sig wallets to a newly generated seed on updated firmware or another secure wallet. Install the latest Coinkite firmware on all devices before generating any new seeds.

When creating seeds, always use at least 50 fair dice rolls plus a strong unique BIP-39 passphrase. Treat any seed created on the listed vulnerable versions without those steps as compromised and retire it.

What happened

Attackers are actively exploiting CVE-2026-18577, an authentication-bypass vulnerability in N-able N-central RMM software, to take over admin accounts and reach managed endpoints.

The flaw is an incomplete patch for the earlier CVE-2026-18556. N-able detected anomalous licensing activity on July 31, 2026; analysis on August 2 revealed the new vector. It affects all versions prior to 2026.3.1.7 (2026.3 Hotfix 1). After admin takeover, attackers abused the Take Control feature, then installed Cloudflare tunnels (service name “Cloudflared”, binary “svchost.exe” in Documents) for persistence. Huntress confirmed ongoing exploitation in customer environments.

Who is affected

Organizations running self-hosted or vendor-hosted N-able N-central (widely used by MSPs) on versions before 2026.3.1.7. Cloud instances receive the hotfix automatically; on-premises deployments must be updated manually.

A limited number of customers have already been breached. Huntress observed >55% of reachable partner/customer cloud servers still unpatched at time of reporting. Downstream impact hits all endpoints managed through the compromised N-central instance.

Why it matters

RMM platforms are high-value targets because they provide privileged, trusted remote access into dozens or hundreds of customer environments at once. A single N-central compromise can cascade into widespread endpoint control, ransomware staging, or data theft across an MSP’s entire client base.

The appliance nature of N-central (custom AlmaLinux, often without EDR) makes detection harder once attackers are inside.

How it could have been prevented

Upgrade all N-central instances to version 2026.3.1.7 or later immediately. For self-hosted servers, apply the hotfix manually and verify success.

Hunt for compromise indicators: admin account anomalies, presence of “svchost.exe” in the Documents folder on managed devices, and a registered service named “Cloudflared”. Review N-able’s published attacker IP list. Rotate credentials and inspect managed endpoints for persistence even after the server is patched.

What happened

A China-linked threat actor using aliases knaithe and KnYuan operated an autonomous attack framework powered by multiple LLMs (primarily DeepSeek via the Hermes Agent, plus Qwen, GLM, Kimi, MiniMax) to scan, select, and exploit internet-facing systems with minimal human input.

Unit 42 discovered the infrastructure after an AI agent misconfigured a file server and exposed logs. The agent received Telegram instructions, then independently queried FOFA for targets, pulled public exploits from GitHub, prioritized by severity and exposure scale, and launched attacks. Documented attempts included CVE-2026-33017 (Langflow, CVSS 9.8) and a chain of CVE-2026-21858 (n8n arbitrary file read, CVSS 10.0) plus CVE-2025-68613 (n8n sandbox escape RCE, CVSS 9.9). No successful compromises were observed in the recovered sessions, but the workflow was fully functional end-to-end.

Who is affected

Operators of internet-exposed instances of targeted software, notably older vulnerable versions of Langflow (prior to 1.9.0) and n8n (e.g., 1.117.3 and other builds before the relevant patches). FOFA showed hundreds of thousands of n8n instances worldwide.

Any organization running unpatched, publicly reachable services that appear in search-engine results used by such agents is in scope; the actor focused on high-severity, high-prevalence targets.

Why it matters

This demonstrates practical, low-supervision AI-driven offensive operations that compress reconnaissance-to-exploit timelines from hours of human work into minutes while self-managing compute. Defenders face adversaries who can continuously re-prioritize the global attack surface faster than traditional patching cycles.

Even failed attempts reveal the maturation of LLM agent tooling for cyber; successful future runs could scale mass exploitation dramatically.

How it could have been prevented

Patch internet-facing applications immediately, prioritizing those with public exploits and high exposure counts (Langflow ≥1.9.0; n8n to versions fixing CVE-2026-21858 and CVE-2025-68613).

Reduce attack surface by removing unnecessary public exposure, requiring authentication on management interfaces, and monitoring for FOFA/Shodan-like scanning followed by exploit attempts. Employ exploit-prevention controls and network segmentation so a single RCE does not yield full environment control.

What happened

Unit 42 researchers detailed novel attack classes against passwordless/passkey authentication, focusing on Google’s synced passkey ecosystem and Cloud Authenticator. Malware on a compromised endpoint can misuse onboarding, recovery, and device-trust workflows to take over passkey-protected accounts.

One key issue is failure by relying parties to properly validate the User Verified (UV) flag. When UV is not enforced, an attacker who already has device access can complete authentication without biometric or PIN user verification, effectively reducing intended multi-factor protection to a single factor. Additional techniques allow authentication without user interaction, bypass of verification requirements, and extraction of synced passkey private keys from the cloud-enclave environment under certain conditions.

Who is affected

Users and organizations relying on passkeys (especially Google’s synced passkey implementation and desktop Cloud Authenticator flows) where the relying party (website or service) does not strictly check the User Verified flag or other WebAuthn assertions.

Any account protected solely by passkeys becomes vulnerable once the endpoint is malware-compromised; scale is potentially billions of accounts as passkey adoption grows.

Why it matters

Passkeys are promoted as the successor that eliminates phishing and shared-secret theft. Implementation gaps and endpoint malware can re-introduce account-takeover paths, undermining the security model just as organizations migrate away from passwords and traditional MFA.

Builders of authentication systems and operators enabling passkeys must treat client-side and RP validation as critical; otherwise the promised phishing resistance and MFA strength evaporate under realistic compromise scenarios.

How it could have been prevented

Relying parties must always validate the User Verified flag and other WebAuthn authenticator data on every assertion; reject authentications that lack UV when MFA is required. Prefer hardware-bound, non-syncable passkeys for high-value accounts where feasible.

On endpoints, enforce strong malware defenses, least privilege, and prompt patching. Monitor for anomalous passkey registration or recovery activity. Users should enable platform biometrics/PIN and avoid approving unexpected passkey prompts.

What happened

CrowdStrike’s 2026 Threat Hunting Report details China-nexus groups Vault Panda (UNC6588) and Genesis Panda (REF0657 / Earth Lamia) exploiting the critical React2Shell vulnerability (unauthenticated RCE in React Server Components and Next.js) within 24 hours of its December 2025 public disclosure and patch release.

The actors rapidly deployed varied malware including RATs for credential harvesting and post-exploitation. Across H1 2026, CrowdStrike found that 88% of publicly disclosed vulnerabilities that were exploited saw intrusion within 48 hours of disclosure. Zero-day exploitation also rose 42% year-over-year from 2024 to 2025. The firm notes these speeds predate widespread frontier-AI use in vuln research, implying further compression ahead.

Who is affected

Organizations running unpatched React Server Components or Next.js applications vulnerable to React2Shell, plus any internet-facing software whose CVEs are publicly disclosed.

Broader impact falls on defenders facing China-affiliated and other rapid-exploitation actors; identity-based attacks and LLMJacking (abuse of stolen corporate LLM API keys) are also rising sharply.

Why it matters

Patch windows have shrunk to hours. Adversaries who monitor disclosures, pre-stage tooling, and validate exploits immediately turn “n-day” vulnerabilities into mass-compromise opportunities before most organizations can react.

AI-assisted vuln discovery and exploitation will accelerate the problem further, raising the cost of delayed patching and incomplete asset inventories.

How it could have been prevented

Maintain a real-time inventory of internet-facing assets and prioritize emergency patching for critical RCEs within hours of disclosure. Subscribe to high-fidelity threat intel and automate virtual patching / WAF rules where available.

Segment networks so initial web-app compromise does not yield credential access or lateral movement. Monitor for rapid scanning and exploitation attempts against newly disclosed CVEs; assume China-nexus and other advanced actors will test your perimeter the same day.

What happened

The UK Police National Legal Database (PNLD) confirmed that contact information for police officers, staff, criminal justice professionals, government partners and customers was compromised and published on the dark web. The incident was identified on 26 July 2026.

Exposed data included names, organisations and work email addresses; some Ask the Police submitter details were also affected. PNLD stated there is no evidence passwords or other credentials were taken. The service is not the Police National Computer or a crime-recording system and holds no victim/witness/offender case data. Researchers linked the broader campaign (ExfilSquad) to misconfigured Microsoft Power Pages / Dataverse portals that allowed anonymous access to tables via Web API or OData.

Who is affected

UK police forces (PNLD supports all 43 Home Office forces and reported ~108,000 police registrations), police staff, criminal justice professionals, government partners, PNLD customers, and some members of the public who used Ask the Police.

Exact victim count has not been disclosed. Contact data of this population is now in criminal hands and circulating on the dark web.

Why it matters

Even “only” names and work emails of law-enforcement personnel enable highly convincing spear-phishing, vishing, and social-engineering attacks that can impersonate colleagues or officials. This raises risks to officer safety, operational security, and public trust.

The apparent root cause—over-permissive anonymous access on a low-code Power Platform portal—highlights how modern SaaS/low-code misconfigurations can expose sensitive directories at scale.

How it could have been prevented

Audit all Power Pages / Power Apps / Dataverse portals for Anonymous Users role permissions on tables; remove public access to any non-public data and disable unnecessary Web API/OData endpoints.

Force password resets and enable phishing-resistant MFA for affected accounts where possible. Train staff on heightened spear-phishing risk using the leaked details. Monitor for spoofed communications claiming to be from PNLD or police colleagues. PNLD has notified the ICO, NCA and affected organisations.

What happened

OpenAI disrupted and banned a coordinated network of ChatGPT accounts assessed to originate from Cambodia’s Preah Sihanouk province, an area linked to online scam compounds and human trafficking. The network used the models to create and manage fake personas, generate/translate scam messages, produce promotional content, and handle internal administrative tasks.

Scam types included investment, romance, gambling, and law-enforcement impersonation. Operators followed a “ping–zing–sting” pattern to build trust then extract money. Some account activity referenced recruitment, debt bondage, detention, and forced criminality consistent with trafficking reporting. A WhatsApp lead earlier in the year triggered the investigation. Scale of losses is unknown but communications referenced hundreds of targets and individual victims losing thousands of dollars.

Who is affected

Potential scam victims worldwide targeted via WhatsApp, Telegram and social media with AI-polished messages and fake profiles. Also the coerced workers inside the compounds whose activities were partly administered via the AI tools.

OpenAI users and the broader ecosystem benefit from removal of the abusive accounts; the action does not directly remediate already-scammed individuals.

Why it matters

Generative AI dramatically lowers the language, scale, and professionalism barriers for scam operations, including those run from trafficking-linked compounds. Platform-level detection and ban waves are necessary but reactive; criminals will migrate to other models or self-hosted tools.

The case underscores both the dual-use risk of capable LLMs and the value of abuse-monitoring combined with external leads (e.g., WhatsApp) for disruption.