Daily Cybersecurity News - April 28, 2026

A critical, unpatched vulnerability identified as CVE-2026-25874 in Hugging Face’s LeRobot open-source robotics platform allows unauthenticated attackers to achieve remote code execution. The flaw stems from unsafe data deserialization over exposed endpoints and currently remains unpatched, posing severe risks to AI inference systems and connected robotics hardware.

• Affected Product: Hugging Face LeRobot versions through 0.5.1
• Identifier: CVE-2026-25874
• CVSS Score: 9.3 (Critical)
• Exploitation Status: Unpatched (Fix planned for v0.6.0)

• Triage: Critical urgency for internet-facing or distributed GPU deployments; isolate the service immediately.
• Attack Vector: Unsafe deserialization via Python’s pickle.loads() over unauthenticated, non-TLS gRPC channels.
• Ease of Exploit: High (Low complexity); requires only network access to the PolicyServer port to send a crafted payload.

• Immediate Action: Restrict network exposure and upgrade to Version 0.6.0 once officially released.
• Workaround: Bind the gRPC server strictly to localhost (127.0.0.1), implement TLS, and enforce authentication.
• Detection: Monitor for anomalous process spawning from LeRobot Python processes and unexpected connections to gRPC ports.

The GlassWorm threat actor has launched a new supply chain attack targeting the OpenVSX ecosystem with 73 “sleeper” extensions. These initially benign clones turn malicious post-update to steal developer credentials and cryptocurrency wallets.

• Initial Access: Supply chain compromise via 73 cloned “sleeper” extensions on the OpenVSX registry.
• Execution: Malicious updates abuse transitive dependencies to fetch loaders and GitHub-hosted VSIX payloads.
• Defense Evasion: Extensions remain benign during marketplace checks, later utilizing invisible Unicode characters and Zig-compiled droppers.

GlassWorm has evolved its tradecraft by using transitive dependencies to bypass marketplace security checks. This shift to delayed execution significantly increases the campaign’s stealth and survivability.

• Target Profile: Software developers using OpenVSX-compatible IDEs (VS Code, Cursor) and AI coding assistants.
• Infrastructure: Leverages the Solana blockchain for primary C2, with Google Calendar as a backup.

• Supply Chain Attack: A cyberattack targeting a software vendor or repository to compromise users who install the trusted software.
• C2 (Command and Control): Infrastructure used by attackers to communicate with and control compromised systems.

An attacker compromised the popular elementary-data PyPI package to distribute an infostealer. The malicious update targeted data engineers and CI/CD pipelines to harvest cloud credentials, SSH keys, and cryptocurrency wallets.

• Initial Access: Exploited a script-injection vulnerability in a GitHub Actions workflow via a malicious pull request.
• Execution: Deployed a base64-encoded payload that executed upon package installation or Docker image pull.
• Defense Evasion: The backdoored release was cryptographically signed and visually identical to a legitimate update.

This supply chain attack highlights the severe risk of CI/CD pipeline hijacking. By compromising a trusted data observability tool, attackers achieved a massive blast radius across enterprise cloud environments.

• Target Profile: Data engineers and enterprise CI/CD pipelines using the Data Build Tool (dbt) ecosystem.
• Infrastructure: Leveraged the official PyPI repository and GitHub Container Registry to distribute the payload.

• Infostealer: Malware designed to covertly gather sensitive information, such as passwords and cryptocurrency wallets, from a compromised system.
• CI/CD Pipeline: Continuous Integration and Continuous Deployment; automated workflows used by developers to build, test, and deploy code.

BlueNoroff, a Lazarus Group subgroup, is executing a massive spear-phishing campaign against Web3 organizations. Attackers use fake Calendly invites and AI-generated Zoom meetings to deploy malware.

• Initial Access: Spear-phishing via manipulated Calendly invites with typo-squatted Zoom links.
• Execution: Fake meeting pages deliver ClickFix clipboard injection, triggering fileless PowerShell execution.
• Defense Evasion: Bypasses Chrome app-bound encryption and injects AES-encrypted shellcode into browser processes.

Attackers use a deepfake pipeline combining stolen webcam footage with GPT-4o portraits to fabricate convincing participants. The execution chain achieves full compromise in under five minutes.

• Target Profile: Web3 and cryptocurrency companies (80%), specifically targeting CEOs and founders (45%).
• Infrastructure: Over 100 victims across 20+ countries using typo-squatted domains.

• ClickFix: A social engineering technique tricking users into copying and executing malicious code via their clipboard.
• Fileless PowerShell: A stealthy attack executing malicious scripts directly in memory to evade traditional antivirus detection.

The DOJ indicted a Chinese national for a multi-year spear-phishing campaign targeting NASA, the US military, and universities. The actor impersonated researchers to steal export-controlled aerospace and weapons development software.

• TTP 1: Spear-phishing via generic email services (Gmail) to impersonate legitimate US researchers.
• TTP 2: Social engineering targets to willingly hand over restricted source code.
• TTP 3: Bypassing export controls by exploiting established trust relationships.

• Threat Actor: Song Wu (Aviation Industry Corporation of China)

• Outcome: Indicted on 14 counts of wire fraud and 14 counts of aggravated identity theft; currently at large.

This campaign highlights how low-tech social engineering can successfully bypass sophisticated perimeter defenses to exfiltrate highly restricted intellectual property.

• Spear-phishing: A targeted attempt to steal sensitive information by masquerading as a trustworthy entity.
• Export Controls: Federal laws restricting the transfer of sensitive technology and software to foreign entities.

Anthropic’s Claude Mythos Preview is a frontier AI model capable of autonomously discovering and weaponizing software vulnerabilities. It fundamentally shifts the cybersecurity landscape by automating complex exploit development against critical infrastructure.

• Autonomous exploitation: Discovers and weaponizes zero-day vulnerabilities across major OSs and browsers.
• Reverse engineering: Reconstructs plausible source code from closed-source binaries to find flaws.
• Advanced chaining: Combines multiple bugs into sophisticated exploit chains to bypass sandboxes.

For Red teams, this drastically lowers the barrier to creating complex, multi-stage exploits. Blue and SOC teams must pivot from periodic assessments to continuous, AI-augmented validation, as the speed of threat discovery now outpaces traditional human-led remediation models.

Gated research preview (Project Glasswing) via AWS Bedrock and Google Vertex AI; strictly invitation-only.

• Zero-day: A software vulnerability unknown to the vendor, meaning no patch currently exists.
• Exploit chain: A sequence of multiple vulnerabilities leveraged together to achieve a broader attack objective.