New GPUThor Rowhammer Attack Bypasses NVIDIA ECC, Enables Root‑Level Escalation on Ampere GPUs
What Happened — Researchers at the University of Toronto disclosed “GPUThor,” a Rowhammer‑style attack that can flip bits on NVIDIA Ampere‑class GPUs fast enough to defeat SECDED ECC protection, cause denial‑of‑service resets, and ultimately achieve root‑level privilege escalation.
Why It Matters for Compliance & Audit Readiness
- Demonstrates a hardware‑level control gap that can invalidate the “integrity” principle of SOC 2 ‑ organizations must prove that critical compute resources are protected against low‑level tampering.
- Continuous evidence of hardware‑security controls (e.g., ECC status, firmware updates, monitoring of abnormal memory‑access patterns) becomes essential audit artefacts.
- Mapping this new attack vector to your control framework helps maintain a defensible audit trail and satisfies the “System Operations” and “Change Management” criteria of SOC 2.
Who Is Affected — AI/ML workloads, cloud service providers, high‑performance computing clusters, and any organization that relies on NVIDIA RTX A4000‑A6000 GPUs for data‑intensive processing.
Recommended Actions
- Verify ECC is enabled on all production GPUs and document the configuration as part of your asset inventory.
- Implement continuous monitoring for abnormal memory‑access patterns and GPU reset events; capture logs as SOC 2 evidence.
- Review firmware and driver versions; apply NVIDIA patches that address known Rowhammer mitigations.
- Update your risk register to include hardware‑level Rowhammer threats and map them to the SOC 2 “System Security” control.
Source: BleepingComputer
Technical Notes
- Attack vector: Rowhammer‑style bit‑flipping (VULNERABILITY_EXPLOIT) that avoids GDDR6 Target Row Refresh (TRR).
- Demonstrated on RTX A4000, A4500, A5000, A6000 (Ampere, GDDR6).
- ECC detects 387 double‑bit errors and mis‑corrects two triple‑bit errors, leading to data corruption and GPU resets every ~2 hours.
- No public CVE assigned yet; the technique is a novel research finding.
Source: same as above