Isometric cyan data tubes exposing Windows Hello PIN bypass risks.

Daily Cybersecurity News – November 18, 2026

CyberRecaps is supported by its readers. We may earn an affiliate commission at no extra cost to you if you buy through a link on this page.

Ghost in the Wires by Kevin Mitnick

TPM Lockout Bypass Cracks Windows Hello PINs

High

What happened

Researchers at Technische Universität Berlin showed that on AMD-based PCs and some Intel PCs, an attacker with physical access who opens the case can reset the TPM lockout counter for Windows Hello and BitLocker PINs.

The attack targets firmware TPMs that store state (including failed-guess counts) encrypted on the mainboard flash chip. By intercepting or glitching the flash (via a small interposer board or voltage glitch on Winbond RPMC chips), the attacker restores a prior state so short PINs can be brute-forced without permanent lockout. In lab tests a 4-digit Hello PIN and a 6-digit BitLocker PIN were recovered in 236 tries.

AMD confirmed a processor firmware flaw enabling the attack on some systems and plans fixes for most chips between November and February 2027.

Who is affected

Users of AMD systems (and certain Intel systems) that rely on firmware TPM for Windows Hello PIN or BitLocker PIN authentication, especially those with short PINs. The tested devices included an HP OmniBook with Ryzen AI 5 and a Framework Laptop 13 with Intel Core Ultra.

Anyone leaving a machine unattended long enough for case opening and flash access is exposed. Scale is significant because fTPM is common on modern laptops and desktops.

Why it matters

Default short PINs become practical to crack in hours or days instead of months or years once the lockout is bypassed. An attacker who already has brief physical access (stolen laptop, evil maid, supply-chain tampering) can unlock the device or BitLocker volume far more easily.

This undermines a core hardware-backed defense that many organizations and individuals rely on for endpoint encryption and passwordless login.

How it could have been prevented

Lengthen Windows Hello and BitLocker PINs by at least two extra digits immediately (researchers' interim advice).

Apply AMD firmware updates as they roll out through early 2027. Prefer discrete TPMs where available, enable Secure Boot and measured boot, and treat unattended physical access as a high risk. Monitor for unexpected TPM state or boot anomalies.

Relevant professional terms

TPM
A Trusted Platform Module is a secure chip (or firmware equivalent) that stores keys and enforces limits such as PIN lockouts to protect device encryption and authentication.
RPMC
Replay-Protected Monotonic Counter is a flash-chip feature that only increments, intended to detect rollback of saved TPM state; the attack exploits incomplete verification or key extraction from these counters.

Java4Shell Enables Pre-Auth RCE in Jira Jenkins

High

What happened

A flaw dubbed Java4Shell (CVE-2026-114938) in Java's accelerated Base64 decoder allows pre-authentication remote code execution on multiple widely used applications when they process malformed input before authentication checks.

The vulnerable fast path activates on x86-64 Intel/AMD CPUs supporting AVX-512 VBMI after JIT compilation. A string ending with an extra '=' causes a fixed-byte write past the buffer end. Attackers arrange memory so the overflow expands a neighboring array (JavaBleed), then gain arbitrary read/write and code execution. It affects Java 17.0.6 and later (and listed Oracle/GraalVM builds). Oracle and OpenJDK released fixes; Amazon Corretto is patched, Red Hat still preparing. CVSS 8.1 (High). No confirmed active exploitation; researchers withheld full exploits and only crashes/memory leaks have been widely reproduced so far.

Default configs of Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate allow unauthenticated exploitation; others require a login.

Who is affected

Organizations running vulnerable Java versions (Oracle Java SE 17/21/25/27 builds noted, GraalVM equivalents, and downstream distributions) on AVX-512 VBMI hardware, especially internet-facing instances of Jira, Jenkins, Keycloak, Confluence, Artifactory, PingFederate, and related apps that decode Base64 early (Basic auth, SAML, CSRF tokens).

Servers behind proxies that fully buffer requests may block some variants, but not all.

Why it matters

Pre-auth RCE on core enterprise tools (CI/CD, identity, collaboration) means unauthenticated attackers can seize servers, steal secrets, or move laterally. The fixed overflow bytes plus memory grooming make reliable exploitation non-trivial yet achievable per target, and failed attempts can crash the JVM for retry.

High CVSS and broad app surface make rapid patching essential even without confirmed in-the-wild use.

How it could have been prevented

Apply the Oracle/OpenJDK (and vendor) Java updates released for the affected builds immediately. Prefer non-vulnerable Java versions or disable the accelerated Base64 path if a temporary workaround exists.

Restrict network exposure of Jira/Jenkins/Keycloak etc., place them behind authenticating proxies that reject incomplete or malformed requests, and monitor for repeated Base64 decode crashes or anomalous memory behavior. Inventory all Java runtimes and pin to patched releases.

Relevant professional terms

Base64 decoder
A routine that turns Base64-encoded text (common in HTTP headers and tokens) back into binary data; the faster SIMD version of this routine contained the overflow.
JavaBleed
The researchers' technique of using the controlled overflow to enlarge an adjacent array, then leveraging that array for arbitrary memory read and write primitives leading to code execution.

Microsoft PyPI Package Poisoned for Cloud Credential Theft

Critical

What happened

Threat actors published two malicious versions (1.26.1 and 1.26.2) of Microsoft's agent-framework meta-package on PyPI on 17 November 2026, 33 minutes apart. Both were later quarantined.

Once installed, each wheel drops a .pth file (agent_framework_meta.pth or af_runtime_diag.pth) into site-packages. Python automatically executes import lines in .pth files on every interpreter start, so the credential stealer runs even if the package is never imported and even during pip itself. The payload harvests environment variables, Azure CLI and other credential files, Azure managed-identity and OIDC tokens, IMDS instance metadata, and more, then zlib-compresses, Base64-encodes and exfiltrates the data to a webhook.site endpoint. It can also request fresh tokens from the platform. How publishing access was obtained remains unclear. Multiple security firms (Aikido, Chainguard, Endor Labs, Socket, Wiz and others) independently reported the incident.

Who is affected

Any Python environment that installed agent-framework 1.26.1 or 1.26.2, including developer workstations, CI/CD pipelines (especially Azure DevOps), Azure App Service/Functions, and containers that pulled the package during the window it was live.

Because the stealer activates on every Python start, secondary tools and scripts in the same environment are also compromised. Microsoft Agent Framework users building AI agents are the primary population.

Why it matters

A trusted Microsoft-branded package became a silent cloud-credential harvester. Attackers obtain long-lived and short-lived Azure, Git, Docker, npm and other secrets plus the ability to mint new tokens, enabling full cloud account takeover, lateral movement and data theft.

.pth-based persistence survives normal package use and is hard to notice without integrity checks. Supply-chain compromises of official packages remain one of the highest-leverage attack paths.

How it could have been prevented

Immediately uninstall versions 1.26.1 and 1.26.2 if present, rotate all credentials and tokens that may have been exposed (Azure, Git, Docker, npm, etc.), and audit Azure activity and IMDS token requests.

Pin dependencies to known-good hashes, enable PyPI 2FA/trusted publishers where possible, scan site-packages for unexpected .pth files, use virtual environments or containers with read-only package installs, and monitor outbound traffic to unfamiliar webhook endpoints. Prefer private package indexes with malware scanning.

Relevant professional terms

PyPI
The Python Package Index is the public repository where developers download Python libraries; a compromised package can reach thousands of projects automatically.
.pth file
A site-packages file whose import lines Python executes automatically at every interpreter startup, giving malware a stealthy, import-independent persistence mechanism.

Operation ShadowCloud Deploys Cloud Backdoors on Governments

High

What happened

Kaspersky researchers identified Operation ShadowCloud, a cyber-espionage campaign active since early 2026 that targets government organizations in Asia and Africa with undocumented modular backdoors.

Initial access often uses the MeshCentral Agent remote-management tool. Attackers then employ DLL loaders to inject encrypted payloads. Two primary backdoors, GitDoor and GraphDoor, provide shellcode execution, plugin download, keylogging, credential theft, screenshots and SOCKS5 proxying. Command-and-control blends into normal traffic by abusing legitimate cloud services including Microsoft SharePoint, GitHub and Ably. Tactical overlaps exist with China-nexus groups such as Bronze Sandcastle, FamousSparrow and CopperTurtle, though attribution remains unconfirmed.

Who is affected

Government sector entities across Asia and Africa. Organizations that allow or use MeshCentral, or that permit unrestricted outbound access to SharePoint, GitHub and similar collaboration platforms, face elevated risk.

Exact victim count is not public, but the campaign's focus on government networks implies sensitive national and diplomatic targets.

Why it matters

Cloud-hosted C2 and modular memory-resident backdoors let operators maintain long-term access while evading traditional network signatures and host-based detection. Governments lose control of credentials, keystrokes and internal traffic, enabling espionage and further staging.

The pattern of living-off-trusted-cloud-services is becoming standard for advanced actors and raises the bar for defenders who must inspect encrypted SaaS traffic and remote-management tools.

How it could have been prevented

Restrict or tightly monitor MeshCentral and similar remote-management agents; require MFA and least privilege. Apply application allow-listing and memory-protection controls to block unauthorized DLL injection and shellcode.

Inspect and constrain outbound connections to SharePoint, GitHub and real-time messaging services; deploy CASB or SSE solutions that detect anomalous C2 patterns. Hunt for the named backdoors and unusual SOCKS proxies. Keep endpoint detection updated and segment government networks.

Relevant professional terms

Backdoor
Hidden malware that gives an attacker ongoing remote control of a system after the initial compromise, often with modular features for spying or lateral movement.
Living-off-trusted-services
An evasion technique in which malware uses legitimate cloud platforms (SharePoint, GitHub, etc.) for command-and-control so that traffic looks like ordinary business activity.
Source: SC Magazine

Google Predicts Autonomous Agentic Attacks Soon

High

What it means

  • Expect machine-speed exploitation of known weaknesses within months.
  • Current attacks are noisy - use that window to tune behavioral detections.
  • Stealth will improve; shift from IOC hunting to anomaly and volume monitoring.
  • Adopt Mandiant/Google hardening guidance and test containment at AI pace.

What happened

Google Threat Intelligence Group (GTIG) stated in its latest quarterly AI Threat Tracker that fully autonomous agentic attacks are likely only three to six months away, following a paradigm shift in the past quarter toward agentic penetration-testing frameworks.

Frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI (plus custom harnesses) orchestrate sub-agents across the full kill chain: reconnaissance, initial access, persistence, privilege escalation, lateral movement and objectives. They discover low-complexity vulnerabilities, generate single-use payloads with LLMs on demand, and adapt from real-time execution feedback. Current operations trade stealth for speed and leave forensic traces, but stealth is expected to improve. The FBI recently highlighted ARTEX abuse, including against South Korean financial institutions. Mandiant released a hardening guide and defensive framework; Google SecOps deployed behavioral/volumetric detections for ARTEX.

Who is affected

Enterprises and governments of all sizes, especially those with unpatched systems, weak credential hygiene or limited behavioral monitoring. Financial services and other high-value sectors already seeing ARTEX activity are early targets.

Any organization whose attack surface can be scanned and exploited at machine speed faces elevated risk as these tools mature.

Why it matters

Autonomous agents compress attack timelines from days to minutes and scale across many victims simultaneously while still using classic techniques (stolen credentials, unpatched vulns). Defenders lose the traditional human-speed advantage.

Even while current tooling is noisy, the trajectory toward quieter, fully agentic intrusions means detection and response must shift to behavioral and volumetric signals rather than known-bad indicators alone.

How it could have been prevented

Deploy behavioral and volumetric detections tuned for agentic frameworks (Google has released ARTEX-focused rulepacks). Follow Mandiant's hardening guide: patch aggressively, enforce credential hygiene and MFA, segment networks, and monitor for unfamiliar tool use or sudden activity spikes.

Assume low-complexity vulnerabilities will be found and exploited automatically; prioritize attack-surface reduction and rapid containment playbooks that do not rely on human-in-the-loop for every step.

Relevant professional terms

Agentic AI
AI systems that can plan, use tools and take multi-step actions toward a goal with limited or no human supervision.
Kill chain
The sequenced stages of a cyber attack from initial reconnaissance through exploitation, persistence, lateral movement and final objectives; agentic frameworks automate progression across these stages.

Docker Hub Keys Stolen in Seconds

High

What happened

Researchers from Stony Brook University published 300 honeypot images containing planted AWS, OpenAI, Anthropic, SSH and database credentials across eight public container registries and measured real-world theft speed.

Images were pulled more than 75,000 times. Planted AWS keys saw over 140,000 API calls from roughly 5,000 IPs. The fastest AWS key abuse occurred 7.86 seconds after publication on Docker Hub; median time to first API call was 51 seconds, and every one of 100 AWS keys was used within 15 minutes. OpenAI and Anthropic keys were rapidly drained. Scanners (collectively called ContainerBots) overwhelmingly extract secrets by static analysis of image layers or registry APIs without ever running the container. Docker Hub dominated the abuse activity. Results were presented at ACM CCS.

Who is affected

Anyone who accidentally or intentionally publishes container images containing cloud API keys, SSH keys, database credentials or similar secrets to public registries, especially Docker Hub.

Development teams, CI systems and organizations that bake secrets into images or leave them in layers are exposed. The scale of automated scanning means virtually every public image is inspected almost immediately.

Why it matters

Secrets in public images have effectively zero safe lifetime. Attackers can validate and weaponize cloud keys before a developer even notices the leak, leading to account takeover, crypto-mining, data exfiltration or further supply-chain abuse.

Static layer inspection means multi-stage builds or .dockerignore mistakes still leak historical secrets. This forces a fundamental change in how container secrets are handled.

How it could have been prevented

Never embed long-lived credentials in container images. Use runtime secret injection (orchestrator secrets, IAM roles, short-lived tokens) instead.

Scan all images for secrets before push (and continuously on registries), enable registry vulnerability and secret scanning, prefer private registries, and treat any public push of a secret-bearing image as an immediate incident requiring key rotation. Monitor cloud API usage for anomalous calls from new IPs.

Relevant professional terms

Container image
A packaged snapshot of an application and its dependencies that can be run as a container; layers can unintentionally contain secrets that scanners extract without execution.
Static analysis of layers
Inspecting the filesystem contents of image layers directly via registry APIs or unpacking, without instantiating the container, to harvest embedded credentials at high speed.

TheMailer Kit Uses AI to Personalize Phishing

Medium

How it works

  1. Operator supplies an AI model or API endpoint and target lists.
  2. Kit generates individualized subjects, bodies and pages from victim context.
  3. Messages are sent at scale while adapting to spam-filter feedback.
  4. Harvested credentials or sessions are collected in the kit's panel.

What happened

A commercial phishing kit known as TheMailer has added attacker-supplied AI model integration that automatically personalizes lures, improves evasion of spam filters, and scales campaigns to tens of thousands of targets.

Operators supply their own AI models or endpoints; the kit uses them to craft individualized email content, subject lines and landing pages based on scraped or purchased victim data. The result is higher-quality, context-aware phishing that bypasses traditional signature and bulk-detection defenses while remaining easy for low-skilled actors to operate.

Who is affected

Enterprises and individuals targeted by phishing campaigns that employ TheMailer or similar AI-augmented kits. Any organization whose users receive external email is potentially exposed, with particular risk to sectors already heavily phished (finance, healthcare, government, tech).

Scale reaches tens of thousands of recipients per campaign, amplifying reach for both cybercrime and targeted operations.

Why it matters

AI personalization collapses the historical quality gap between mass phishing and spear-phishing. Defenders can no longer rely on generic lures looking obviously fake or on simple volume-based filters.

Commercial availability lowers the skill barrier, so more actors can run high-volume, high-believability campaigns. This increases successful credential harvests, malware delivery and business-email-compromise risk.

Relevant professional terms

Phishing kit
A ready-made toolkit that lets attackers build, host and send fake login pages and emails without writing the infrastructure from scratch.
AI-personalized lure
Email or landing-page content dynamically generated by a language model using victim-specific data so each message appears uniquely relevant and trustworthy.
Source: Dark Reading