# SentinelLabs Benchmark: Frontier LLMs Attempt Autonomous Long-Horizon Malware Analysis Using the 2005 Pre-Stuxnet 'fast16' Sabotage Toolkit as Case Study

> SentinelLabs ran an eight-stage benchmark testing whether frontier AI models (GPT-5.6 Sol, GPT-5.5, GLM-5.2, Claude Opus 4.6/4.7/4.8, Grok 4.2, Gemini 3.1 Pro, DeepSeek Reasoner) can autonomously conduct a long-horizon reverse-engineering investigation of the 2005 Windows sabotage toolkit 'fast16' -- a pre-Stuxnet precision-sabotage implant that targeted nuclear-program engineering software. Only GPT-5.6 Sol completed all eight stages; researchers conclude frontier models are best used as supervised investigative agents rather than autonomous replacements for senior reverse engineers, citing recurring semantic errors and premature promotion of weak findings even in the strongest runs.

- **Published:** 2026-07-22T00:00:00Z
- **Last reviewed:** 2026-11-06T12:04:11.391Z
- **Canonical:** https://intel.threadlinqs.com/threat/TL-2026-1630
- **ID:** TL-2026-1630
- **Severity:** INFORMATIONAL
- **Category:** THREAT_INTEL
- **Status:** ACTIVE
- **Detections:** 9 (1 shown free below) · **IOCs:** 31 (3 shown free below)

## Description

In April 2026, SentinelLabs researchers Vitaly Kamluk and Juan Andres Guerrero-Saade (with contributions from Silas Cutler and Costin Raiu) publicly disclosed 'fast16' at Black Hat Asia: a previously unknown Windows sabotage malware framework compiled in mid-to-late 2005, predating the public disclosure of Stuxnet by roughly five years. The discovery originated from a carrier executable, svcmgmt.exe, uploaded to VirusTotal in 2016 but bearing an August 2005 PE compile timestamp. svcmgmt.exe operates as a Windows service host that, depending on command-line arguments, can run as a service, execute an embedded Lua 5.0 interpreter (extended with a custom Unicode 'wstring' module and a built-in symmetric cipher routine for decrypting payload blocks), or interpret a filename argument to spawn two further commands. A companion boot-start kernel-mode driver, fast16.sys, implements a patch engine of 101 byte-pattern rules that locate specific code sequences inside targeted 32-bit Windows executables (Intel-compiler-linked .EXE files), capture original addresses, splice in replacement code -- including an injected floating-point-unit (FPU) block for 'precision arithmetic and scaling values in internal arrays' -- and repair PE metadata/checksums to avoid detection. This is not a rootkit in the conventional sense; SentinelLabs stresses the driver's function is highly targeted binary patching of computational logic, not file-hiding.

The likely targets identified via static/dynamic analysis and corpus testing are engineering/scientific simulation packages: LS-DYNA 970 (an explosion/impact/multiphysics finite-element solver used in structural and munitions modeling), PKPM (a Chinese structural-engineering CAD suite with the SATWE analysis engine), and MOHID (a Portuguese hydrodynamic/environmental modeling platform) -- the latter two sharing Fortran compiler lineage with LS-DYNA and triggering only 'haphazard' rule collisions rather than meaningful patches. The Institute for Science and International Security has documented Iran's historical use of LS-DYNA within its nuclear weapons-related engineering programs -- the same broader program later targeted overtly by Stuxnet -- which SentinelLabs cites as circumstantial evidence that fast16 may have been an early, covert attempt to sabotage nuclear-relevant physics simulations by introducing silent, systematic numerical errors rather than exfiltrating data or causing an overt crash. A reference to fast16-like tradecraft ('*** Nothing to see here - carry on ***') was found in the April 2017 Shadow Brokers 'Territorial Dispute' leak's drv_list.txt, a file believed to catalog NSA operator deconfliction/avoid-fratricide signatures, further supporting a nation-state origin. Source-control artifacts (SCCS/RCS style markers) embedded in the code point to developers working in long-term government- or military-grade Unix development environments.

Building on this real-world case, SentinelLabs designed an eight-stage benchmark to test whether current frontier LLMs could perform the equivalent long-horizon reverse-engineering investigation autonomously, end to end, without an operator doing the analytical work. Stages progressed from producing a gold-master IDA Pro IDB and identifying components (Stage 1), through patch-logic impact assessment on a single LS-DYNA installer (Stage 2), generalizing across a wider LS-DYNA corpus (Stage 3), correctly rejecting unsuitable candidates PKPM and MOHID (Stage 4), forming and testing hypotheses against unavailable ANSYS AUTODYN targets (Stage 5), validating claims against external public reporting (Stage 6), producing a publication-grade report (Stage 7), and reassessing conclusions under newly introduced ANSYS AUTODYN evidence (Stage 8). Models were graded on the quality of the underlying IDA Pro 9.3 databases they produced, not on prose summaries, and evaluated for local analytical soundness plus whether that quality held as investigative scope expanded ('investigative horizon').

Only one model configuration -- GPT-5.6 Sol, across three runs at different reasoning-effort levels -- completed all eight stages. Sol's distinguishing behavior was responding to local failures with structural process fixes: when a mistake slipped through, Sol would reassess its own quality-control process and add scaffolding (e.g., codifying file hashes for routine operations) to prevent recurrence, and in one xhigh-effort run correctly walked back an over-broad hypothesis about rule applicability to PKPM/MOHID after evidence showed only incidental fragment matches, while preserving the valid LS-DYNA findings. GPT-5.5 ran extended autonomous sessions but never escaped an early 'Level 1' grind phase and failed at result aggregation. GLM-5.2, run under a hard $100 / 267.41M-token budget, showed capable local experimentation but suffered a severe setback recovering from 113 lost functions and only reached Level-1-equivalent performance. Claude Opus 4.6/4.7/4.8 runs showed capable error correction and evidence-responsive revision but repeatedly over-promoted milestones before defects were resolved and left the investigation incomplete. Earlier runs with Grok 4.2 (reasoning mode), Gemini 3.1 Pro, and DeepSeek Reasoner all converged on a shared misdiagnosis, misidentifying fast16 as a conventional filesystem rootkit rather than a targeted binary patcher. Aggregate token consumption across all models/effort levels exceeded 23 billion tokens, with a 93.58% prompt-cache read rate.

SentinelLabs' overarching conclusion is that even the strongest frontier model (Sol) still made serious semantic errors, accepted weak internal quality controls, promoted work before verification, and depended on human-operator review at consequential decision points -- meaning current frontier LLMs should be deployed as supervised investigative agents (a force multiplier under expert oversight) rather than autonomous replacements for senior malware reverse engineers. The report frames this as directly relevant to closing the decades-long reverse-engineering expertise bottleneck in the security industry, while cautioning that the same emerging capability trajectory has dual-use implications for adversary-side automation of malware analysis and weaponization research.

## MITRE ATT&CK

- T1543 Create or Modify System Process
- T1547 Boot or Logon Autostart Execution
- T1036 Masquerading
- T1027 Obfuscated Files or Information
- T1140 Deobfuscate/Decode Files or Information
- T1685 Disable or Modify Tools
- T1070 Indicator Removal
- T1518 Software Discovery
- T1083 File and Directory Discovery
- T1082 System Information Discovery
- T1059 Command and Scripting Interpreter
- T1106 Native API
- T1565 Data Manipulation
- T1485 Data Destruction
- T1005 Data from Local System
- T1559 Inter-Process Communication
- T1587 Develop Capabilities
- T1569.002 System Services
- T1078.001 Valid Accounts
- T1110.001 Brute Force
- T1021.002 Remote Services
- T1570 Lateral Tool Transfer
- T1135 Network Share Discovery
- T1480 Execution Guardrails
- T1112 Modify Registry

## Free detection (Splunk SPL, Microsoft KQL, Sigma)

**fast16-style Service Registration Followed by Mass Security-Product Enumeration** (`TL-2026-1630-SPL-001`)

Detects a Windows service registered under a generic system-management name (e.g. SvcMgmt) followed within a short window by rapid sequential registry queries against many distinct AV/HIPS vendor service keys from the same process — the fast16 carrier's pre-execution 18-product security-software discovery sweep before deciding whether to proceed.

- **Languages:** Splunk SPL, Microsoft KQL, Sigma
- **ATT&CK:** T1543, T1518, T1036
- **Data sources:** windows_security_eventlog, windows_system_eventlog, sysmon
- **Severity:** high
- **Confidence:** medium

### Splunk SPL

```spl
index=windows sourcetype="WinEventLog:System" EventCode=7045
| eval svc_name=lower(Service_Name)
| where match(svc_name, "svcmgmt|sysmgmt|winmgmtsvc") OR match(lower(Service_File_Name), "cmd\.exe\s*/c\s*exit")
| rename Service_File_Name as image_path
| join type=left host max=0 [
    search index=windows sourcetype="WinEventLog:Security" EventCode=4663 OR EventCode=13
    | regex Object_Name="SYSTEM\\\\CurrentControlSet\\\\Services\\\\(Symantec|Sygate|TrendMicro|ZoneLabs|FSecure|Kaspersky|McAfee|Panda|Sophos|ESET|Avira|BitDefender|Malwarebytes|Comodo|GData|QuickHeal|ClamWin)"
    | stats count as av_probe_count dc(Object_Name) as distinct_av_keys by host
]
| where av_probe_count >= 6
| table _time host Service_Name image_path av_probe_count distinct_av_keys
| sort -_time
```

### Microsoft KQL

```kql
let svcCreation = DeviceEvents
| where Timestamp > ago(24h)
| where ActionType == "ServiceInstalled"
| where AdditionalFields has_any ("SvcMgmt", "SysMgmt", "WinMgmtSvc");
let avProbing = DeviceRegistryEvents
| where Timestamp > ago(24h)
| where RegistryKey has @"SYSTEM\CurrentControlSet\Services"
| where RegistryKey has_any ("Symantec","Sygate","TrendMicro","ZoneLabs","FSecure","Kaspersky","McAfee","Panda","Sophos","ESET","Avira","BitDefender","Malwarebytes","Comodo","GData","QuickHeal","ClamWin")
| summarize probeCount = count(), distinctKeys = dcount(RegistryKey) by DeviceId, InitiatingProcessId, bin(Timestamp, 5m);
svcCreation
| join kind=inner (avProbing) on DeviceId
| where distinctKeys >= 6
| project Timestamp, DeviceId, probeCount, distinctKeys, AdditionalFields
```

### Sigma

```yaml
title: fast16-style Service Registration with Mass AV/Security-Product Enumeration
id: 7c1a9f2e-4b3d-4e1a-9a6f-1d2c3e4f5a6b
status: experimental
description: Detects a generically-named Windows service creation event followed by rapid sequential registry enumeration of multiple distinct AV/HIPS vendor service keys, consistent with fast16 carrier (svcmgmt.exe) pre-execution security-software discovery tradecraft.
references:
    - https://www.sentinelone.com/labs/fast16-mystery-shadowbrokers-reference-reveals-high-precision-software-sabotage-5-years-before-stuxnet/
author: Threadlinqs Team
date: 2026-07-22
tags:
    - attack.persistence
    - attack.t1543.003
    - attack.discovery
    - attack.t1518
    - attack.tool.fast16
logsource:
    product: windows
    category: process_creation
detection:
    selection_svc:
        EventID: 7045
        ServiceName|contains:
            - 'SvcMgmt'
            - 'SysMgmt'
    selection_probe:
        EventID: 4663
        ObjectName|contains:
            - 'Symantec'
            - 'Sygate'
            - 'TrendMicro'
            - 'ZoneLabs'
            - 'FSecure'
            - 'Kaspersky'
            - 'McAfee'
            - 'Panda'
    timeframe: 60s
    condition: selection_svc and selection_probe | count(ObjectName) by Image >= 6
falsepositives:
    - Legitimate AV/EDR deployment or inventory-management tooling enumerating security-product registry keys during onboarding
level: high
```

**Known false positives:**

- Legitimate AV/EDR management or inventory tools that enumerate security-product service keys as part of onboarding or compliance scanning

**Tuning notes:** Tune the av_probe_count threshold to the number of distinct AV vendor keys realistic in your environment; exclude known asset-inventory/EDR-deployment service accounts from the join to reduce false positives.

### 8 more detection rules (Splunk SPL, Microsoft KQL, Sigma), unlock with TL Creds

- svcmgmt-named Service Installation with Security Product Registry Sweep (locked: unlock with TL Creds)
- Unsigned Boot-Start Driver Load with Subsequent PE Checksum-Field Rewrite on Third-Party Binaries (locked: unlock with TL Creds)
- [indicator] Parent Process Spawning Embedded-Interpreter Decode-then-Command-Dispatch Chain (locked: unlock with TL Creds)
- fast16-style Service Creation with Security-Product Enumeration Follow-up (Sigma) (locked: unlock with TL Creds)
- Boot-Start Kernel Driver Registration Followed by Rapid Multi-Process Binary Modification (Sigma) (locked: unlock with TL Creds)
- Named-Pipe IPC Consistent with fast16 p577-style Component Signaling (Sigma) (locked: unlock with TL Creds)
- Boot-Start Unsigned Kernel Driver Followed by Third-Party Executable Byte-Range Modification (locked: unlock with TL Creds)
- Short Numeric-Suffix Named Pipe IPC Preceding Symmetric-Cipher Decode-then-Dispatch (locked: unlock with TL Creds)

Sign up free to unlock the detection rules: https://threadlinqs.com/?modal=signup

## Sample indicators (3 of 31)

Defanged, TLP:CLEAR.

- File hash: `9a10e1faa86a5d39417cae44da5adf38824dfb9a16432e34df766aa1dc9e3525`
- File hash: `dbe51eabebf9d4ef9581ef99844a2944`
- File hash: `07c69fc33271cf5a2ce03ac1fed7a3b16357aec093c5bf9ef61fbfa4348d0529`

Full set of 31 indicators unlocks with TL Creds: https://threadlinqs.com/tl-creds

## Sources

- [Frontier Models Tackle Autonomous Long-Horizon Malware Analysis](https://www.sentinelone.com/labs/frontier-models-tackle-autonomous-long-horizon-malware-analysis/)
- [fast16 | Mystery Shadow Brokers Reference Reveals High-Precision Software Sabotage 5 Years Before Stuxnet](https://www.sentinelone.com/labs/fast16-mystery-shadowbrokers-reference-reveals-high-precision-software-sabotage-5-years-before-stuxnet/)
- [Pre-Stuxnet Sabotage Malware 'Fast16' Linked to US-Iran Cyber Tensions](https://www.securityweek.com/pre-stuxnet-sabotage-malware-fast16-linked-to-us-iran-cyber-tensions/)
- [Researchers find sabotage malware that may predate Stuxnet](https://www.theregister.com/2026/04/24/fast16_sabotage_malware/)
- [Fast16 malware existed before the Stuxnet worm appeared](https://hackmag.com/news/fast16)
- [Fast16: Pre-Stuxnet Sabotage Tool Was Built to Subvert Nuclear Weapons Simulations](https://www.security.com/threat-intelligence/fast16-nuclear-sabotage)
- [Fast16 Malware Targets High-Value Systems With Sabotage Capabilities](https://gbhackers.com/fast16-malware/)
- [Fast16: The Pre-Stuxnet Cyber Weapon That Rewrote History](https://jonathansblog.co.uk/fast16-the-pre-stuxnet-cyber-weapon-that-rewrote-history)
- [Fast16: Cyber Sabotage Before Stuxnet](https://arnav.au/2026/04/26/fast16-cyber-sabotage-before-stuxnet/)
- [Previewing GPT-5.6 Sol: a next-generation model](https://openai.com/index/previewing-gpt-5-6-sol/)
- [GPT-5.6: Frontier intelligence that scales with your ambition](https://openai.com/index/gpt-5-6/)

## Full data

The other 8 detection rules (Splunk SPL / Microsoft KQL / Sigma) unlock with TL Creds; a free account includes daily creds: https://threadlinqs.com/?modal=signup
Indicator values unlock with TL Creds, at a lower price on Red and Purple; AI agents can query rules and IOCs through the Threadlinqs MCP server (Purple): https://intel.threadlinqs.com/mcp

Canonical: https://intel.threadlinqs.com/threat/TL-2026-1630
