Anthropic Details Linux CVE-2026-39962 Exploit; Cursor Pulls Backdoor

Anthropic's Frontier Red Team published a technical analysis on Friday, Oct. 30 detailing how its frontier model, Claude Mythos Preview, autonomously discovered and constructed an unauthenticated remote code execution exploit for CVE-2026-39962, a CVSS 9.8 race condition in the Linux kernel's Multipath TCP (MPTCP) networking subsystem (Anthropic Frontier Red Team).

Introduced when MPTCP was merged into the mainline Linux kernel in March 2020 and patched in Linux 6.19.8 on March 13, 2026 after Anthropic reported the flaw to kernel maintainers on March 7, CVE-2026-39962 occurs when a subflow socket resets while an MPTCP listener is still completing a connection acceptance. Given five hours and access to two virtual machines without internet access, Claude Mythos Preview turned a 16-byte null write primitive into remote root execution without an information leak. The model accomplished this by spraying msg_msg structures across the SLUB allocator to construct a self-referencing linked list, using a blind memory copy gadget to forge a pipe_buffer structure pointing into kernel code despite Kernel Address Space Layout Randomization (KASLR), and redirecting socket close operations through the extended Berkeley Packet Filter (eBPF) subsystem's __bpf_prog_Run interpreter to execute shellcode as root.

Microsoft Analyzes Persistent SSH Backdoor in Tampered Cursor Extension

In a separate software supply-chain disclosure on Friday, Microsoft Threat Intelligence published a forensic breakdown of tampered versions 2.1.312 and 2.1.313 of the anthropic.claude-code extension that were served in the Cursor extension marketplace on Oct. 29 between 2:27 AM and 1:35 PM PDT (Microsoft Security Blog). Both Microsoft and Anthropic confirmed that neither the official Visual Studio Marketplace nor the npm package @anthropic-ai/claude-code was compromised.

According to Microsoft's analysis, the injected JavaScript loader executed prior to normal extension initialization, collected host details and up to 40 environment-variable names matching keywords such as KEY, TOKEN, SECRET, PASS, AWS, or GCP, appended an attacker-controlled Ed25519 public key labeled svc-telemetry to ~/.ssh/authorized_keys, and initiated an outbound reverse SSH tunnel (-N -R '*:0:localhost:22') to 134.209.82.225 while installing platform-specific startup persistence under names such as com.apple.sysmond.plist on macOS, a @reboot crontab entry on Linux, and a SysMon scheduled task on Windows.

Microsoft Teams External Moderation, Apple Spyware Alerts, and Pixel 10 Hardening

Across enterprise and mobile security on Friday:

• Microsoft Teams Cross-Tenant Reporting: Microsoft introduced a public preview in Teams allowing users to report suspicious messages from external tenants and quarantine flagged threads for administrator review, while rolling out AI agents in Security Copilot to detect brand impersonation following campaigns by threat groups including Storm-2603 (which used Teams voice calls to deploy Black Basta and LockBit ransomware) and Storm-1811 (The Record by Recorded Future).
• Apple Mercenary Spyware Notifications: Apple issued a new wave of threat notifications on Oct. 28 to iPhone users in more than 100 (100+) countries, with Dutch Member of Parliament Jesse Klaver confirming on Oct. 30 that he received an alert warning his device had been targeted by mercenary spyware (SecurityWeek).
• Google Pixel 10 Hardware Defenses: Google engineers outlined hardware and firmware protections in the Pixel 10 series powered by the Tensor G5 chip, combining Arm Memory Tagging Extension (MTE), Rust-rewritten media and DNS parsers, cellular 2G disablement, and a hardened Advanced Protection Mode that deactivates JavaScript JIT compilation and restricts USB data connections while the screen is locked (Help Net Security).