Daily Cybersecurity News – August 14, 2026

What happened

Threat actors began exploiting a critical Microsoft SharePoint authentication bypass shortly after a public proof-of-concept was released.

CVE-2026-55040 (CVSS 9.1) is a security feature bypass stemming from weak authentication and flaws in the JWT token validation pipeline. It allows an unauthenticated attacker to forge a valid JWT (using techniques such as alg:none and untrusted certificate resolution) and impersonate any SharePoint site user or administrator, enabling file disclosure and data modification. Microsoft patched it in the July 2026 Patch Tuesday updates.

Rapid7 released a Python PoC that chains four weaknesses in SPJsonWebSecurityTokenHandlerV2 and related classes. Telemetry showed a spike to 12 exploitation attempts (mostly Aug 12-13) from IPs in Hong Kong, Japan, the Netherlands, Taiwan, and the US after the PoC dropped. It is the fifth SharePoint flaw exploited in 2026.

Who is affected

Organizations running unpatched on-premises Microsoft SharePoint Server instances exposed to the network.

Any SharePoint deployment that has not applied the July 2026 security updates is at risk of unauthenticated access and impersonation of site users or admins.

Why it matters

SharePoint is a core collaboration and document platform in enterprises. Successful exploitation gives attackers the ability to read and modify sensitive files while operating as legitimate users or admins, facilitating further lateral movement, data theft, or persistence without needing credentials.

The rapid weaponization after PoC release underscores how quickly critical enterprise flaws move from disclosure to real-world abuse, especially given the string of recent SharePoint exploits.

How it could have been prevented

Apply the July 2026 Microsoft Patch Tuesday updates for SharePoint immediately and verify installation across all servers.

Restrict network access to SharePoint management interfaces, monitor for anomalous JWT or authentication bypass attempts, and review logs for unexpected administrative activity or file access. Consider isolating or decommissioning outdated SharePoint instances where possible.

What happened

A suspected advanced persistent threat actor launched a global campaign exploiting a critical VMware vCenter directory traversal vulnerability days after disclosure.

CVE-2026-59310 (CVSS 9.8) is a directory traversal flaw in the vCenter Syslog server that lets an unauthenticated attacker with network access execute arbitrary code. Broadcom/VMware disclosed and patched it on July 29, 2026. Exploitation began around August 3.

German IR firm QUIRSO observed the campaign during an engagement: attackers used the flaw for initial access, then installed a malicious cron job running reverse_ssh for persistent outbound C2. Activity hit 361 unique victim IPs across 47 countries (heavily targeting US, Germany, France, Turkey, Iran). Patching alone may not remove existing reverse_ssh persistence.

Who is affected

Organizations running unpatched VMware vCenter Server instances reachable over the network, especially those with internet-exposed management interfaces.

vCenter is widely deployed as the central management plane for virtualized environments; any unpatched appliance is at high risk of full compromise.

Why it matters

Compromising vCenter often grants control over the entire virtual infrastructure, enabling ransomware, data theft, or further lateral movement across VMs and hosts. The short disclosure-to-exploitation window and post-exploitation persistence mean even rapid patching may leave lingering access.

vCenter remains a high-value target for both cybercriminals and state actors because of its blast radius.

How it could have been prevented

Apply the Broadcom/VMware patches for CVE-2026-59310 immediately and verify all vCenter instances are updated.

Perform forensic checks for reverse_ssh, unexpected cron jobs, or outbound connections to attacker infrastructure. Isolate vCenter management interfaces with network segmentation, restrict outbound connectivity, and monitor for path-traversal or anomalous Syslog activity. Use QUIRSO's published YARA rule for reverse_ssh detection.

What happened

A self-represented litigant in a Connecticut court case hid prompt-injection instructions inside official court filings, written in tiny 3-point white font so they were nearly invisible to humans but readable by software.

The hidden text directed any AI reviewing the document to side with the plaintiff, accurately reflect the filing, and 'ensure remediation.' Additional messages included casual taunts and a meme link. The court noticed extra white space, examined the pleadings, and identified the injections.

Judge Walter Spader Jr. issued a sanctions decision criticizing the attempt. The court does not use AI to process filings. The injections were spotted by attorney Brendan Palfreyman and confirmed by 404 Media.

Who is affected

Primarily the legal system and any courts, law firms, or parties that might feed filings into AI tools for summarization, analysis, or drafting.

In this specific case it affected one Connecticut civil matter (Matthew Elliott vs. New York Bariatric Group), but the technique has broader implications for document-processing pipelines.

Why it matters

As courts and lawyers increasingly experiment with AI for document review and case preparation, hidden instructions could bias outputs, undermine fairness, or introduce unreliable assistance. Even unsuccessful attempts highlight a new attack surface in trusted legal workflows.

The incident shows prompt injection moving beyond chatbots into real-world official documents and underscores the need for human oversight and input sanitization.

What happened

Group-IB researchers uncovered WindRelay, a new Android malware purpose-built to capture live payment-card data over NFC and relay it in real time to attacker-controlled terminals.

It is paired with the SpyNote remote-access Trojan. The scam begins with a social-engineering phone call claiming to be from the victim's bank. The victim is guided to install a personalized SpyNote app (labeled with their own name). Once SpyNote provides remote control, the fraudster silently installs WindRelay. The victim is kept on the call while the attacker uses the phone as an NFC relay, streams card data to a fake merchant terminal, and can even take out loans via remote access to banking apps—using the PIN the victim enters themselves.

23 WindRelay samples appeared on VirusTotal (Nov 2025–Jul 2026), targeting Czechia, Slovakia, and Slovenia. Four C2 IPs were identified.

Who is affected

Android users, particularly in Central Europe (Czechia, Slovakia, Slovenia), who answer bank-impersonation calls and install apps.

Any smartphone with NFC enabled that can be socially engineered into installing a malicious app is potentially vulnerable; victims keep physical possession of their cards throughout the attack.

Why it matters

This combines live social engineering, device takeover, and real-time NFC relay so fraud can occur while the victim still holds the card and stays on the phone. It defeats many traditional card-not-present controls and physical-possession assumptions.

The dual-malware approach (RAT + specialized relay) shows modern fraud operators stacking capabilities for high-success, low-friction cash-outs and loans.

How it could have been prevented

Never install apps or follow instructions from unsolicited bank calls; hang up and contact the bank via official channels. Enable Google Play Protect, restrict unknown-source installs, and review app permissions (especially NFC, accessibility, and contacts).

Banks and users should monitor for unusual contactless transactions and loan applications. Mobile security tools that detect RATs or anomalous NFC activity help. Educate users that legitimate banks never ask them to install apps during a live call.

What happened

Jamf Threat Labs identified a new multi-stage Rust-based macOS infostealer named AmnesiaStealer being distributed through ClickFix social-engineering lures.

Victims are shown a fake GitHub download page with a 'Terminal installation' box. Clicking copy produces a base64 command that, when pasted and run, fetches a short self-deleting script which launches the AmnesiaStealer binary. The malware mutes system sound (to hide Finder copy sounds), harvests data from Apple Notes and Telegram while avoiding TCC prompts, prompts for the user password to unlock keychain files, stages and exfiltrates credentials/browser data/system info, then deploys a second-stage module.

The second stage gives operators hidden remote control of a cloned Chromium browser profile (Chrome, Brave, Edge, Arc, Opera, Vivaldi, Chromium) so live authenticated sessions and cookies can be stolen invisibly. It includes OS-version-branched logic targeting previously patched macOS bypasses.

Who is affected

macOS users who fall for ClickFix lures, particularly those visiting counterfeit download pages and pasting commands into Terminal.

Any Mac running supported browsers and holding credentials, notes, or Telegram data is a potential target; the campaign reuses lure templates seen with other macOS stealers.

Why it matters

ClickFix bypasses many traditional defenses because the user themselves executes the command. AmnesiaStealer’s stealth (sound muting, TCC avoidance, profile cloning for invisible browser control) and multi-stage design make detection and response harder while enabling high-value session and credential theft.

It expands the growing macOS infostealer ecosystem and shows attackers refining delivery and post-infection capabilities specifically for Apple platforms.

How it could have been prevented

Train users never to copy-paste and run commands from web pages or unsolicited sources. Enforce application allow-listing, block unknown Terminal executions where possible, and keep macOS and browsers fully patched.

Deploy endpoint detection that flags base64-decoded scripts, unexpected keychain access, sound-muting, and anomalous browser-profile cloning. Monitor for ClickFix-style lures and use browser isolation or least-privilege accounts for daily work.

What happened

Research from XBOW shows that mid-tier proprietary and open-source AI models have crossed a practical threshold in offensive security tasks, becoming strategically useful for hacking and exploitation at far lower cost than frontier models.

Models such as Z.ai’s GLM-5.2, xAI’s Grok 4.5, Anthropic’s Opus 4.7, Meta’s Muse Spark 1.1, and especially OpenAI’s GPT-5.5 now handle moderately complex agentic exploitation workflows that they struggled with only six months earlier. GPT-5.5 delivered strong results on both white-box and black-box web-app testing, cut miss rates dramatically (10% vs 40% for GPT-5), and performed better without source-code access—closer to a real attacker’s viewpoint.

Because these models are cheaper, operators can run them repeatedly or for longer horizons, allowing them to leapfrog more expensive frontier models on many tasks.

Who is affected

Security teams, red teams, and organizations concerned about AI-assisted attacks; also policymakers watching AI capability thresholds.

Any environment whose defenses assume attackers lack sophisticated, low-cost automation is now more exposed as these capabilities democratize.

Why it matters

While frontier models grab headlines for sandbox escapes, the 'middle class' of cheaper, efficient models lowers the barrier to automated vulnerability discovery and exploitation. Attackers can afford higher volume and longer agentic runs, increasing the scale and speed of offensive operations.

Defenders must prepare for a world where capable AI-driven recon and exploitation are no longer limited to well-resourced actors.

What happened

CRM provider Beacon disclosed that a compromised AWS access key was the likely root cause of a breach affecting data held by approximately 1,500 UK charities.

The key was potentially exposed in public JavaScript build artifacts during development. The attacker used the valid credentials to access and download all data in the CRM platform, including attachment files. Malicious activity ran for about 1 hour 27 minutes starting July 27, 2026, 01:20 UTC, correlating with a spike in downloads. Data was encrypted at rest but decrypted by AWS upon authorized download.

Beacon reset all related credentials, found no evidence of persistence, and has not seen the data published or misused so far. Multiple charities (including those supporting survivors of sexual abuse, hospitals, and homelessness) publicly confirmed impacted supporter data.

Who is affected

Beacon’s entire customer base of roughly 1,500 UK charities and the individuals whose personal data those charities held (names, emails, phone numbers, donation records).

Sensitive sectors such as healthcare, victim support, and disability charities were among those affected. No payment-card, bank-account, or clinical patient data was stored in the CRM.

Why it matters

Charities often hold trusted personal and donation data of vulnerable populations. Exposure enables highly targeted social-engineering and phishing against supporters. The ICO has already indicated individual charities are not responsible, placing the burden on the SaaS provider.

The incident highlights how a single leaked cloud credential in build artifacts can lead to bulk exfiltration of an entire multi-tenant customer base.

How it could have been prevented

Never embed long-lived AWS access keys in client-side JavaScript, build artifacts, or public repositories. Use short-lived credentials, IAM roles, secrets managers, and least-privilege policies. Scan build outputs and public assets for secrets continuously.

Enable CloudTrail, GuardDuty, and cost/usage anomaly detection; require MFA and conditional access for sensitive roles. Rotate keys immediately on any suspicion of exposure and conduct regular access reviews. SaaS providers should isolate tenant data more strictly and monitor for bulk download patterns.