AMD RCCL EUVD-2026-93987

Exploitation requires the attacker to control a peer rank within an RCCL (ROCm Communication Collectives Library) collective-communication group - i.e., participation in a multi-GPU/multi-node distributed workload on AMD Instinct hardware - or a network-adjacent position on the collective fabric; this maps to PR:L, so some authenticated/privileged foothold in the compute environment is needed, not anonymous internet access. The flaw is an improper input validation leading to an attacker-controlled pointer dereference (CWE-822). Limiting factors: AC:H indicates specific conditions/timing or data shaping are required, RCCL is only exercised in distributed GPU workloads (not a default internet-facing service), and the attack surface is confined to AMD Instinct MI200/MI300-class GPU deployments running the affected RCCL version.

This is a moderate-priority issue rather than a drop-everything emergency, and the signals are largely consistent. The CVSS 4.0 vector (AV:N/AC:H/AT:N/PR:L/UI:N) describes a network-reachable but high-complexity flaw that requires some level of privilege - consistent with the description's 'compromised peer rank or network-adjacent attacker,' meaning the attacker must already be a participant in, or adjacent to, an RCCL collective-communication group rather than an arbitrary internet attacker. Technical impact is high (VC:H/VI:H/VA:H), and the untrusted-pointer-dereference weakness (CWE-822) plausibly supports remote code execution. However, no EPSS score, no CISA KEV listing, and no public POC were provided, so there is no public-exploitation signal. The combination of AC:H (high attack complexity), PR:L (requires privilege), and a deployment footprint limited to AMD Instinct datacenter GPU clusters strongly argues for targeted, insider-or-tenant-level risk rather than mass exploitation. Net: a genuine high-impact bug worth patching in AI/HPC environments, but not a high-probability widespread-exploitation threat on current evidence. No conflicting signals were observed, but the absence of EPSS/KEV/POC means exploitation likelihood cannot be quantified from the input.

In a multi-GPU AMD Instinct training or HPC cluster, an attacker who controls one peer rank (e.g., a malicious tenant, compromised node, or network-adjacent host on the collective fabric) sends crafted data through RCCL that fails input validation and causes an attacker-controlled pointer to be dereferenced. Under the right conditions this can corrupt memory and lead to arbitrary code execution on the targeted node. No public exploit code was identified at time of analysis, and the high attack complexity means exploitation is non-trivial.