A vulnerability (CVE-2026-4747) has been addressed in FreeBSD that allows code execution at the kernel level through sending network packets to the NFS server. The issue occurs when using the kgssapi.ko module that implements the RPCSEC_GSS API on the kernel side. Besides the kernel, the vulnerability also affects user space applications utilizing the librpcgss_sec library and functioning as an RPC server. Such applications, which are not part of the core FreeBSD system, could also be attacked by sending network packets.
The issue exists in the implementation of the GSS (Generic Security Services) API, which facilitates the establishment of secure authenticated communication channels with proxy server. RPCSEC_GSS is used in NFS servers to protect access to Sun RPC through Kerberos-based authentication and to encrypt traffic between the server and client. The problem arises because, during signature verification, data from the packet is copied into a fixed buffer without proper size checks. The flaw appears before authentication is completed but requires the ability to send network packets to the NFS server. An exploit is available that allows remote execution of /bin/sh with root privileges.
The vulnerability was discovered by an employee of Anthropic using the AI assistant Claude. Notably, external researchers from the Calif team used Claude to write a working exploit, providing only the publicly released general report on the vulnerability from FreeBSD. In addition to directly exploiting the vulnerability, the AI model deployed a virtual machine with a vulnerable configuration, set up remote debugging and kernel crash dump reading, and organized an exploit to run /bin/sh after achieving code execution at the kernel level. The exploit took 4 hours of Claude's processing time to create.
The researchers who created the exploit did not stop there and continued experimenting, using Claude to identify vulnerabilities in Vim and Emacs that allow for code execution when opening specially crafted files in the editors. Notably, the prompts provided to the model were reduced to a primitive task statement, such as "find a 0-day vulnerability in Vim that occurs when opening a file". As a result, the Claude model successfully discovered previously unknown vulnerabilities.
The vulnerability in Vim (CVE-2026-34714) is caused by an error in processing the tabpanel option in the default modeline mode (":set modeline"), which allows for the specification of editing options in the processed file. According to the developers' design, only a limited number of options can be set through the modeline, with the expressions executed in sandbox mode that only permits the simplest safe operations.
The tabpanel option did not have the P_MLE flag set, allowing the use of the expression %{expr}, which executes without activating the modelineexpr mode. A flaw in the autocmd_add() function was exploited, where proper security checks were missing when binding an action to the SafeStateAgain event, allowing the command to be executed after exiting the sandbox isolation. This vulnerability was fixed in the release of Vim v9.2.0272. An example line that triggers the execution of the 'id' utility and directs the output to the file '\/tmp\/calif-vim-rce-poc': /* vim: set showtabpanel=2 tabpanel=%{%autocmd_add([{‘event’\:’SafeStateAgain’,’pattern’\:’*’,’cmd’\:’!id>\/tmp\/calif-vim-rce-poc’,’once’\:1}])%}: *\/
The vulnerability in Emacs is caused by the automatic processing of the contents of the .git\/ directory when it is located in the same directory as the file being opened. In this case, Emacs executes the commands "git ls-files" and "git status" within the context of the " .git\/" contents when opening the file. To execute code, it is enough to open a file in Emacs from a directory that contains a .git\/ subdirectory with a configuration file "config" that includes the attacker's command for execution in the option "core.fsmonitor". The maintainers of GNU Emacs refused to address the vulnerability, stating that the issue lies within Git.
Additionally, two more vulnerabilities can be noted:
- CVE-2026-33150 — a use-after-free memory access in the io_uring handler from the libfuse library, potentially allowing code execution upon resource exhaustion while working with filesystems implemented through FUSE (Filesystem in Userspace).
- CVE-2026-34743 — a buffer overflow in the lzma_index_append() function implementation from the liblzma library. The issue occurs when using the lzma_index_decoder() function to decode an index without records, resulting in allocation of a buffer smaller than required. It is noted that the low-level lzma_index* API is rarely used in applications, making it unlikely to find applications that meet the exploitation conditions when working with the index. The vulnerability has been fixed in the XZ Utils 5.8.3 update.
Source: opennet.ru
