libriscv CVE-2026-62958

The attacker must be able to supply a crafted ELF binary that libriscv/rvlinux will load - i.e. a deployment that parses attacker-controlled binaries, which is libriscv's intended sandbox use case. The specific trigger is an ELF whose .text or .symtab section header has a manipulated sh_offset (and/or sh_size), e.g. set to a large value such as 0x7FFFFFFFF0000000 in the PoC; these section-header fields bypass the PT_LOAD program-header bounds checks that binary_loader performs, because serialize_execute_segment and lookup() index m_binary with sh_offset without range-checking it. Affected versions are libriscv up to and including 1.17. Limiting factors: the attack is local (AV:L) in that it requires getting a binary into the emulator, so deployments that only run trusted locally-built binaries are not realistically exposed; the demonstrated outcome is a host crash, while host memory disclosure additionally requires favorable heap layout or integer overflow (grooming) that the advisory does not demonstrate.

This is a genuine but bounded risk whose priority depends entirely on deployment model. The source-provided CVSS 3.1 vector (AV:L/AC:L/PR:N/UI:N/S:C/C:L/I:N/A:H, 7.9 HIGH) correctly captures the shape: local attack vector (the attacker must feed a binary to the emulator), low complexity, a scope change (guest/ELF input compromising the host emulator process), confirmed high availability impact (reliable host crash) and a low, only-theoretical confidentiality impact. The vuln.today decision label is this_week, not act_now/emergency, which is consistent with the absence of a KEV listing and the absence of any EPSS score in the input. No EPSS value was provided, so exploitation-probability modeling cannot be compared here; no SSVC exploitation/automatable/technical-impact breakdown was provided either, only the aggregate priority label. A public proof-of-concept (poc.py) is published in the GHSA advisory and is demonstrated to crash rvlinux with SIGSEGV, so working trigger code for the denial-of-service case is in the open. The key signal to weigh carefully: the confidentiality dimension (C:L, i.e. host memory disclosure) is described by the advisory as possible 'through grooming' and is explicitly NOT demonstrated - only the host SIGSEGV is reproduced. Real priority is therefore high specifically for anyone running libriscv <=1.17 as a multi-tenant or untrusted-binary sandbox (exactly its intended use), and low for deployments that only ever load trusted, locally-produced binaries, since AV:L means an attacker needs a path to submit a crafted ELF.

An attacker who can submit an ELF binary to a service built on libriscv (a sandbox-as-a-service that runs user-supplied RISC-V programs) uploads a binary whose .text and/or .symtab section headers carry a malicious sh_offset. When the emulator parses the binary this offset is added directly to the host buffer pointer and dereferenced out of bounds, instantly crashing the host process (demonstrated SIGSEGV via the published poc.py). With a favorable heap layout or integer overflow the same flaw could instead map host heap memory into the guest, but this disclosure path is described as possible rather than demonstrated.