🕵️ RESEARCH & DEEP DIVES
Critical PwnEscape flaw (CVE-2026-99999) in Linux kernel networking lets guest VMs read host memory
CVE-2026-99999
Researchers disclosed PwnEscape, a critical Linux kernel networking flaw enabling guest-to-host memory leaks and lab-proven VM escape, demonstrated against Vercel's production sandboxes.- Affects any Linux host running VMs on shared kernel networking code — including cloud sandbox platforms like Vercel's Firecracker microVMs — not a specific hypervisor
- CVE-2026-99999 (CVSS 9.3, Critical): crafted network packets from inside a guest read ~32 bytes of host memory per packet, crash the host, and execute code with root access
- Host code execution was shown only in lab tests; the memory leak was confirmed by Vercel against its production sandboxes during its $1M challenge
- Researcher Paulos Yibelo (pwn.ai) found the flaw in three days using an AI-assisted research system, costing under $700 in API tokens
- Fixed by kernel commit 6785011f8b16, backported to stable branches: 5.10.272, 5.15.223, 6.1.190, 6.6.159, 6.12.113, 6.18.57, 7.2.11; first fixed in kernel 5.10.272
Unit 42 tracks AI-powered espionage campaign targeting government Linux infrastructure
Palo Alto Unit 42 identified an AI-powered espionage campaign, tracked as CL-UNK-1211, targeting government Linux infrastructure from October 2025 to April 2026.- Targets government entities and state-affiliated infrastructure running enterprise Linux and database environments, tracked by Unit 42 as CL-UNK-1211 (affiliation undetermined).
- Attackers dumped sensitive databases, exfiltrated government mailboxes to access linked social media accounts, and harvested credentials via a hijacked .bashrc sudo alias and malicious PHP code.
- Initial access used custom and open-source web shells, with persistence via modified auto_prepend_file PHP settings under /etc/php/ and /opt/plesk/php/ and a web shell planted at /usr/lib/php/plesk_sys_driver.php.
- Payloads contained AI-generated conversational comments and reasoning notes, suggesting code was produced by LLMs or an automated AI agent, with heavy Base64 use across PHP, JavaScript, Bash and Python.
- Defense evasion included disabling endpoint protections, erasing system logs, and planting deceptive red-team markers to disguise activity as a legitimate security assessment.
📋 ADVISORIES
- 📄 Source for Hackers exploit Collabora CODE RCE flaw (CVE-2026-110447) to backdoor Nextcloud servers — github.com