Threat reportThreat IntelligenceTL-2026-1598
Bit2Watt: Synchronized GPU Power-Oscillation Attack Could Let Cloud Tenants Destabilize Power Grids
Researchers Zhouhao Ji, Kaikai Pan, and Wenyuan Xu (Zhejiang University) disclosed 'Bit2Watt' (CHES 2026), a cyber-physical technique in which a malicious cloud tenant modulates ordinary GPU compute load — via a custom CUDA kernel (SWMA) or manipulated LLM training hyperparameters (LTMA) — to create 1.2-6 kHz power-draw oscillations that evade 1 Hz PDU and 450 Hz NVML telemetry.
Key facts
| Identifiers | TL-2026-1598 |
|---|---|
| Also known as | Watt2Bit |
| Severity | HIGH |
| Affected products | NVIDIA RTX 4090; NVIDIA A100; NVIDIA Tesla V100; Generic Data-center power distribution units (PDUs), UPS, voltage regulator modules, switching power supplies; Generic Grid-connected distributed energy resource (DER) inverters in converter-dominated grids |
| Malware | None (no malware payload used) |
| Tools | Custom CUDA kernel (SWMA proof-of-concept), Impedance-based analysis framework, Near-field EMI antenna |
| MITRE ATT&CK techniques | T1005 Data from Local System, T1020 Automated Exfiltration, T1036 Masquerading, T1052 Exfiltration Over Physical Medium, T1078 Valid Accounts, T1082 System Information Discovery, T1119 Automated Collection, T1129 Shared Modules, T1204 User Execution, T1489 Service Stop, T1495 Firmware Corruption, T1496 Resource Hijacking and 8 more in the attack flow |
| Motivation | UNKNOWN |
| Target sectors | cloud computing, data centers, energy, critical infrastructure, artificial intelligence machine learning |
| Target regions | Global |
| Detection rules | 9 (1 shown free) |
| Sources | 8 cited sources |
Detection coverage
As of 2026-07-21, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1598 across Splunk SPL, Microsoft KQL and Sigma, covering 17 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.
Free detection: SWMA Synthetic Workload Modulation — High-Frequency CUDA Saturation/Idle Toggling
Detects CUDA-referencing processes/telemetry that toggle a GPU node between full-saturation and near-idle compute states with clock-like periodicity, consistent with the Bit2Watt Synthetic Workload Modulation Attack (SWMA) using a cudaMallocManaged-based switching signal to drive 1.5-6 kHz power oscillation invisible to 1 Hz PDU and 450 Hz NVML sampling.
- Languages: Splunk SPL, Microsoft KQL, Sigma
- ATT&CK: T1587.001, T1496, T1082
- Data sources: nvidia_nvml_telemetry, cuda_process_events, hypervisor_gpu_metrics
- Severity: high
- Confidence: medium
Splunk SPL
index=gpu_telemetry sourcetype="nvidia:nvml:metrics" OR sourcetype="cuda:process:events"
| eval util_state=if(gpu_utilization>=95,"SATURATED",if(gpu_utilization<=5,"NEAR_IDLE","MID"))
| where util_state!="MID"
| bin _time span=1s
| stats count(eval(util_state=="SATURATED")) as sat_transitions count(eval(util_state=="NEAR_IDLE")) as idle_transitions dc(util_state) as state_variety values(process_name) as process_name by host, _time
| streamstats window=60 count as toggle_window by host
| where toggle_window>100
| table _time host process_name sat_transitions idle_transitions toggle_window
| sort - toggle_windowMicrosoft KQL
let LookbackWindow = 1h;
DeviceProcessEvents
| where Timestamp > ago(LookbackWindow)
| where ProcessCommandLine has "cuda" or FileName has_any ("cuda", "nvcc")
| join kind=inner (
DeviceEvents
| where Timestamp > ago(LookbackWindow)
| where ActionType == "GpuUtilizationSample"
) on DeviceId
| summarize ToggleCount = count(), SaturatedSamples = countif(AdditionalFields has "utilization=100"), IdleSamples = countif(AdditionalFields has "utilization=0") by DeviceId, bin(Timestamp, 1m)
| where ToggleCount > 100 and SaturatedSamples > 0 and IdleSamples > 0
| project Timestamp, DeviceId, ToggleCount, SaturatedSamples, IdleSamplesSigma
title: Periodic High-Frequency GPU Utilization Toggling (SWMA-consistent)
id: b2w-swma-spl001-2026-1598
status: experimental
description: Detects process/telemetry patterns consistent with the Bit2Watt Synthetic Workload Modulation Attack (SWMA) toggling a GPU between saturation and near-idle states at sub-second, high-repetition intervals not explainable by normal workload variance.
references:
- https://arxiv.org/abs/2607.05993
- https://thehackernews.com/2026/07/new-bit2watt-attack-could-let-cloud.html
author: Threadlinqs Team
date: 2026-07-21
tags:
- attack.resource-development
- attack.t1587.001
- attack.impact
- attack.t1496
- attack.tool.custom_cuda_kernel
logsource:
category: process_creation
product: linux
detection:
selection:
Image|contains: 'cuda'
timeframe: 1m
condition: selection | count() by ComputerName > 100
falsepositives:
- Legitimate bursty inference workloads with irregular (non-periodic) timing
- Autoscaling GPU schedulers with rapid legitimate saturation/idle cycling
level: mediumKnown false positives:
- Legitimate bursty inference or autoscaling workloads with irregular, non-periodic saturation/idle cycling
Tuning notes: Tune the toggle_window/count(100) threshold per rack baseline; sub-1Hz PDU/450Hz NVML sampling won't natively see kHz oscillation, so this rule depends on deploying sub-kHz telemetry per the research remediation before it can fire meaningfully — treat count thresholds as a proxy signal from higher-resolution counters, not the raw 1Hz/450Hz feeds alone.
- Sub-kHz Power-Quality Degradation — Rising THD / Negative Damping Ratio · unlock with TL Creds
- Anomalous Cloud-Tenant GPU Resource-Hijacking Pattern Preceding Power Modulation · unlock with TL Creds
- GPU Telemetry Impair-Defenses Pattern — Modulation Timed Below Sampling-Rate Blind Spot · unlock with TL Creds
- Custom CUDA Kernel Reconnaissance and Capability Development Preceding SWMA · unlock with TL Creds
- Endpoint Denial-of-Service via Data-Center Power/Thermal Protection Trips (Watt2Bit Feedback) · unlock with TL Creds
- FSK-Encoded Near-Field EMI / Power-Line Signal Consistent with Watt2Bit Covert Exfiltration · unlock with TL Creds
- LTMA Hyperparameter/Batch-Size Oscillation Masquerading as Training Noise · unlock with TL Creds
- Coordinated Multi-Tenant GPU Power-Draw Synchronization (Grid-Impact Precondition) · unlock with TL Creds
Indicators of compromise
As of 2026-07-21, Threadlinqs Intelligence tracks 17 indicators of compromise for TL-2026-1598.
Technical analysis
Simulated at 1,000 synchronized GPUs on a 1 MW, 90%-DER-penetration grid, the attack drove current THD to 46.8% (vs. the 13% IEC 61000-3-12 guideline) and pushed the damping ratio negative (-0.27), risking oscillatory instability and up to 81% cascading load-shed blackouts; a reverse 'Watt2Bit' path also enables EMI-based covert data exfiltration.
Bit2Watt is a cyber-physical vulnerability, not a software exploit or CVE-bearing bug: it requires only legitimate, unprivileged GPU compute access inside a multi-tenant cloud or colocation environment. The adversary needs no malware, no stolen credentials, and no vendor-specific flaw — the attack surface is the physical coupling between GPU power draw and the electrical grid supplying the data center.
Two delivery methods are documented. The Synthetic Workload Modulation Attack (SWMA) uploads a custom CUDA kernel that toggles the GPU between high-intensity compute and near-idle states on a host-controlled schedule (using a `cudaMallocManaged` unified-memory flag as the switching signal), achieving power-modulation frequencies from roughly 1.5 kHz up to 6 kHz, peaking on an RTX 4090. The LLM Training Modulation Attack (LTMA) is the more dangerous variant: it embeds the same modulation logic inside an otherwise-legitimate LLM training job by adjusting batch size, auxiliary operations, or hyperparameters, producing lower-frequency (approximately 1.2-3 kHz) but higher-amplitude oscillations that closely resemble normal training noise and are far harder to flag as anomalous.
Standard data-center telemetry cannot see this signal: rack PDU counters sample once per second and NVIDIA's NVML telemetry samples at roughly 450 Hz, both far below the kHz-range modulation frequencies used by the attack. When many GPUs across a shared power domain are synchronized, the aggregate load behaves as a constant-power load with negative incremental resistance, exciting resonant modes in distributed-energy-resource (DER) inverter control loops rather than relying on classical synchronous-generator inertia (converter-dominated grid). In the researchers' simulated worst case — 1,000 synchronized GPUs on a 1 MW local grid at 90% DER penetration — current THD reached 46.8% (versus the 13% IEC 61000-3-12 stability guideline) and the system damping ratio went negative (-0.27), signaling onset of oscillatory instability; a companion wide-area/European grid simulation showed cascading protection trips producing up to 81% load shedding across 13 stages. Locally, the induced harmonics also stress voltage-regulator modules, UPS units, PDUs, and switching power supplies inside the data center itself, wasting substantial current as non-productive heat (~20% excess) and creating a feedback denial-of-service risk (over-temperature/over-current protection trips) back onto the AI cluster — termed the 'Watt2Bit' feedback path.
The same physical channel runs in reverse as a covert side channel: encoding a logical '1' as a 2 kHz power tone and a '0' as 200 Hz, the researchers used a near-field EMI antenna to recover a 50-bit test sequence with zero errors, demonstrating that GPU power modulation can be used for both cyber-physical disruption and covert exfiltration from air-gapped or isolated compute environments.
The attack's principal real-world constraint is tight timing synchronization across many independently scheduled, physically distributed GPUs/tenants — the paper itself concedes this remains 'an open problem,' and introducing 100 microsecond (SD) timing jitter reduced achieved modulation amplitude by roughly 20%. No CVE has been assigned and no vendor patch is applicable; this is an architectural exposure from coupling volatile, unauthenticated multi-tenant GPU compute loads to inverter-heavy, low-inertia electrical grids, not a fixable code defect. It follows a related August 2025 Microsoft/OpenAI/NVIDIA paper that warned synchronized AI training power swings could physically damage grid equipment when their frequency aligns with utility critical frequencies.
MITRE ATT&CK techniques by tactic:
Collection
T1005 Data from Local System; T1119 Automated Collection
Exfiltration
T1020 Automated Exfiltration; T1052 Exfiltration Over Physical Medium
Defense Evasion
Initial Access
Discovery
T1082 System Information Discovery; T1526 Cloud Service Discovery
Execution
T1129 Shared Modules; T1204 User Execution
Impact
T1489 Service Stop; T1495 Firmware Corruption; T1496 Resource Hijacking; T1499 Endpoint Denial of Service; T1529 System Shutdown/Reboot
Command and Control
Resource Development
T1583 Acquire Infrastructure; T1587 Develop Capabilities
Reconnaissance
T1592 Gather Victim Host Information
defense-impairment
Remediation
Immediate actions
- Deploy sub-kHz-resolution power telemetry (beyond 1 Hz PDU / 450 Hz NVML sampling) on GPU racks to detect abnormal modulation signatures
- Monitor for coordinated, synchronized GPU utilization oscillations across tenants sharing a power domain
- Flag CUDA workloads that repeatedly toggle between saturation and near-idle states with regular periodicity
- Baseline LLM training job power/utilization noise profiles to catch LTMA-style hyperparameter-driven modulation anomalies
Workarounds
- Rate-limit or cap per-tenant GPU power-state transition frequency where cloud provider policy allows
- Segregate large multi-tenant GPU clusters across independent power domains/UPS zones to limit worst-case synchronized-tenant count per grid interconnect
Longer-term hardening
- Deploy local energy buffering (batteries, supercapacitors) to absorb kHz-range demand transients before they reach the grid interconnect
- Add harmonic filtering / active power-quality conditioning at data-center points of common coupling
- Redesign DER inverter control loops with cross-layer awareness of aggregated compute-driven load dynamics
- Develop cross-layer (workload-scheduling + power-electronics) defense standards for AI data centers, as recommended by the CHES 2026 authors
- Shield or physically isolate high-value compute racks against near-field EMI monitoring to mitigate the Watt2Bit covert-channel risk
Affected and fixed versions
- NVIDIA — RTX 4090
Vulnerable versions: all - NVIDIA — A100
Vulnerable versions: all - NVIDIA — Tesla V100
Vulnerable versions: all - Generic — Data-center power distribution units (PDUs), UPS, voltage regulator modules, switching power supplies
Vulnerable versions: standard telemetry sampling rates (1 Hz PDU / 450 Hz NVML) - Generic — Grid-connected distributed energy resource (DER) inverters in converter-dominated grids
Vulnerable versions: high DER penetration (~90%) grid segments
Timeline
- Microsoft, OpenAI, and NVIDIA researchers publish a prior paper warning that synchronized AI-training power swings can physically damage grid equipment when their frequency aligns with utility critical frequencies, establishing the research lineage Bit2Watt builds on.
- A companion wide-area/European grid simulation models cascading protection-relay trips across 13 stages, producing up to 81% load-shedding in the worst-case scenario, distinct from the localized 1 MW/1,000-GPU THD test.
- Researchers validate the power-modulation channel on real GPU hardware and grid-connected photovoltaic (PV) inverters using impedance-based analysis frameworks, in addition to pure power-system simulation, showing GPU loads can reach modulation frequencies exceeding 6,000 Hz versus a few hertz for conventional household loads.
- Paper is accepted for publication at CHES 2026 (IACR Conference on Cryptographic Hardware and Embedded Systems).
- Zhouhao Ji, Kaikai Pan, and Wenyuan Xu (Zhejiang University) submit 'Bit2Watt: A Cyber-Physical Vulnerability Exploiting GPU Workloads Across Power and Computing Infrastructures' to arXiv (2607.05993).
- The Register publishes coverage of the Bit2Watt findings, summarizing the grid-destabilization risk from malicious cloud tenants.
- Additional outlets (news-pravda.com, GuardianMSSP) republish or syndicate the Bit2Watt disclosure the same day, extending coverage beyond the initial The Hacker News and Register reporting.
- TL-Intel-Harness ingests the disclosure via RSS hunt (The Hacker News feed) as TL-2026-1598 for detection-engineering awareness despite theoretical/research-only status.
- No CVE is assigned and no vendor (NVIDIA or cloud providers) issues a patch or advisory, as the researchers characterize the exposure as architectural rather than a fixable software defect.
- The Hacker News, GBHackers, TechRadar, GuardianMSSP, and other outlets publish broad coverage of Bit2Watt and its Watt2Bit covert-channel companion finding, driving the hunt-source article for this threat record.
Sources cited
- Bit2Watt: A Cyber-Physical Vulnerability Exploiting GPU Workloads Across Power and Computing Infrastructures arxiv.org · Research · 2026-07-07
- Bit2Watt: A Cyber-Physical Vulnerability Exploiting GPU Workloads Across Power and Computing Infrastructures (HTML) arxiv.org · Research · 2026-07-07
- Bit2Watt Attack Turns AI Data Centers Into Cyber-Physical Threats to Local Power Grids gbhackers.com · Blog · 2026-07-21
- New Bit2Watt Attack Could Let Cloud Tenants Disrupt Power Grids Without an Exploit guardianmssp.com · Blog · 2026-07-21
- New Bit2Watt Attack Could Let Cloud Tenants Disrupt Power Grids Without an Exploit The Hacker News · News · 2026-07-21
- Malicious cloud customers can bring down the power grid The Register · News · 2026-07-20
- Experts warn hackers could shut down entire power grids by hijacking cloud accounts techradar.com · News · 2026-07-21
- Bit2Watt frames GPU workloads as a grid attack surface news-pravda.com · News · 2026-07-21
Previous and next in threat intel
Drafted by the Threadlinqs research pipeline from 8 sources; published ; last updated —.