Detection queries Not classified

ESET GuardBreaker - UAC-0099 plants a refusal-bait comment in a MATCHBOIL script

ESET Research described GuardBreaker in posts on X on Aug 27, 2026. Tomáš Foltýn's WeLiveSecurity article followed on Sep 10, 2026. In a VBScript used early in an attack on a target in Ukraine, UAC-0099…

What happened

ESET Research described GuardBreaker in posts on X on Aug 27, 2026. Tomáš Foltýn's WeLiveSecurity article followed on Sep 10, 2026. In a VBScript used early in an attack on a target in Ukraine, UAC-0099 left inert REM comments. Help Net Security quotes the opening as a request to make a nuclear weapon and to help, followed by uranium-enrichment text. The comment has no runtime effect. ESET says it is there so an LLM-powered code scanner's safety guardrails refuse the file and stop reading before the payload logic. ESET names this a reverse jailbreak: the point is to cause a refusal, not to talk a model into complying. The script's job is to download and install MATCHBOIL, a loader ESET says is used only by UAC-0099. The same group has checked for IDA and Wireshark processes in other attacks. ESET did not test whether the comment works and did not name a targeted model. DugganUSA, in an independent write-up dated Sep 1, 2026, ran one A/B pass on a synthetic inert script: GPT-4o and Mistral analyzed the treatment and flagged the decoy, and Claude Sonnet 4.5 returned an empty refusal. That test is not ESET's, and DugganUSA says it is a single sample on one day. UAC-0099 is Russia-aligned, active since at least 2022, and has targeted Ukrainian transportation and energy. Reporting says the group has handed validated access to Sandworm. On Jul 21, 2026 CERT-UA documented an updated chain: a VBS script with a fake extension and a long run of spaces before .vbs, Notepad++ 8.8.3 components, and the tools LUNCHPOKE, BURNYBEAR, and MATCHBOIL.V2. That advisory is about the group and the loader, not about the GuardBreaker comment. ATT&CK does not have a technique for a refusal bait in a comment. The script execution itself is T1059.005.

The target is an LLM-powered code scanner. An inert comment is meant to trip safety guardrails so the model refuses and stops before the payload logic. ESET did not name a product or test the effect.

Read the original sourceAdditional source: Russian hackers plant a nuclear-weapon prompt in malware to trip AI safety filtersAdditional source: UAC-0099: LUNCHPOKE, BURNYBEAR, updated MATCHBOIL.V2, and Notepad++ 8.8.3Additional source: Independent A/B test of the decoy comment on three models

How it happened

Select a stage to explore its defenses
Detect this stage
Detect this stage
Detect this stage
Detect this stage
Select a stage · Drag to explore · + / − to zoom · Dashed edges are loopsSelect a stage · Arrows show reported connectionsReported sequence
Read the flow as text & sources

A catalog reconstruction of the supplied source account. Unknown steps remain explicit; an edge describes the reported relationship between steps.

  1. 01 · VBScript in an early stage

    UAC-0099 uses a VBScript early in an intrusion against a target in Ukraine. The script's runtime job is to download and install MATCHBOIL.

    Hunt: A script host or shell runs encoded content, a remote URL, or inline script syntax, then downloads or launches a payload.

    Encoded commands and remote administration can be legitimate. This does not cover every script host, in-process execution, or non-Windows branch.

    Source for step 1
  2. 02 · Inert refusal-bait comment

    REM comments request help with a nuclear weapon and then add enrichment text. They do not execute. ESET says they are meant to make an LLM scanner refuse before it reaches the loader logic.

    Hunt: Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.

    This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.

    Source for step 2
  3. 03 · MATCHBOIL still installs

    If the script is run, it still downloads MATCHBOIL, the group's loader. CERT-UA's July chain uses a double-extension VBS, Notepad++ components, LUNCHPOKE, BURNYBEAR, and MATCHBOIL.V2.

    Hunt: A program delivered by the reported lure executes, establishes access or contacts infrastructure outside the approved software baseline.

    A program name alone is not a signature. This normalized template requires the relevant platform collector and does not assert Windows-only policy coverage.

    Source for step 3
  4. 04 · A refusal is the gap

    ESET did not measure the effect. DugganUSA's separate one-day test saw one of three models return an empty refusal and the other two flag the decoy and continue.

    Hunt: Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.

    This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.

    Source for step 4

Why it matters

Reported entry point: A VBScript used in the early stage of an intrusion at a Ukrainian organization. ESET does not describe how the script arrived. CERT-UA's July chain for the same group starts from a phishing ZIP that contains a VBS file disguised with a fake extension. Reported impact: ESET did not measure the effect. DugganUSA's separate one-day test saw one of three models return an empty refusal and the other two flag the decoy and continue; reported targets include An unnamed organization in Ukraine. UAC-0099's usual targets include transportation and energy; applicability to your setup has not been assessed, so verify affected versions, permissions, and configuration in the original report.

Read the original report

What you can do

Protection & prevention

Keep the scanner's instructions separate from the file under analysis, and treat text inside the sample as data. A refusal, an empty completion, or a safety block on a malware sample should raise a review, not end it. Fall back to analysis that does not depend on a model, and do not give one model the only vote on whether a file is safe. The comment does not change what the script does when Windows runs it, so script-host and loader detections still apply.

Catalog defensive guidance based on the reported behavior. Confirm scope and current mitigations in the original source.

Review the original guidance Additional source 2 Additional source 3 Additional source 4
Investigate each phase

Detection queries

Draft query starters grouped by the attack stages where they apply.

Download stage pack

Start with the logs you collect. Each hunt identifies its required visibility, field mapping and limits. Queries are catalog-authored drafts using reusable behavior profiles; test with known activity and your normal baseline before alerting.

01VBScript in an early stageSuspicious Windows script executionNative logs · Draft

UAC-0099 uses a VBScript early in an intrusion against a target in Ukraine. The script's runtime job is to download and install MATCHBOIL.

What to look for

A script host or shell runs encoded content, a remote URL, or inline script syntax, then downloads or launches a payload.

wscript or cscript running the group's VBS. How this script arrived is not described.

Visibility: Defender-side logs where collected

Required logs

  • Windows process creation: Sysmon Event 1 or Defender DeviceProcessEvents
  • PowerShell script-block logging when enabled
  • Download / network and installer events

Coverage limit. Encoded commands and remote administration can be legitimate. This does not cover every script host, in-process execution, or non-Windows branch.

Windows process candidates. Sigma uses normalized process_creation fields; KQL requires Defender for Endpoint; SPL requires the Sysmon XML sourcetype. Tune collectors, fields and baselines before use.

Sigma
Draft SigmaDownload Sigma
title: Hunt - Suspicious Windows script execution
id: 2bb88802-9f03-4d99-a192-fe44c9484e8b
status: experimental
description: A script host or shell runs encoded content, a remote URL, or
  inline script syntax, then downloads or launches a payload.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
  - Authorized activity matching this broad hunt; baseline and investigate
    before alerting.
level: low
logsource:
  category: process_creation
  product: windows
detection:
  selection:
    Image|endswith:
      - \powershell.exe
      - \pwsh.exe
      - \mshta.exe
      - \wscript.exe
      - \cscript.exe
    CommandLine|contains:
      - -enc
      - http
      - FromBase64String
      - "javascript:"
      - "vbscript:"
  condition: selection
KQL
Draft KQLDownload KQL
// Draft hunt; requires Defender for Endpoint process telemetry for this OS.
DeviceProcessEvents
| where Timestamp > ago(7d)
| where FileName in~ ("powershell.exe", "pwsh.exe", "mshta.exe", "wscript.exe", "cscript.exe")
| where ProcessCommandLine contains "-enc" or ProcessCommandLine contains "http" or ProcessCommandLine contains "FromBase64String" or ProcessCommandLine contains "javascript:" or ProcessCommandLine contains "vbscript:"
| project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine, InitiatingProcessFileName, InitiatingProcessCommandLine, SHA1
| order by Timestamp desc
SPL
Draft SPLDownload SPL
sourcetype="XmlWinEventLog:Microsoft-Windows-Sysmon/Operational" EventCode=1
Image IN ("*\\powershell.exe", "*\\pwsh.exe", "*\\mshta.exe", "*\\wscript.exe", "*\\cscript.exe")
CommandLine IN ("*-enc*", "*http*", "*FromBase64String*", "*javascript:*", "*vbscript:*")
| table _time host User Image CommandLine ParentImage ParentCommandLine
| sort - _time
Investigation steps
  1. Follow the parent and child process tree and recover the script or URL from logs without executing it.
  2. Correlate the payload's signer/hash, persistence, and network activity; baseline approved automation.
02Inert refusal-bait commentIndependent malware-analysis validationCustom schema · Draft

REM comments request help with a nuclear weapon and then add enrichment text. They do not execute. ESET says they are meant to make an LLM scanner refuse before it reaches the loader logic.

What to look for

Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.

The comment is inert and is meant to make an LLM code scanner refuse. ESET did not test a product. A refusal on a malware sample should start a review.

Visibility: Review / research workflow

Required logs

  • Sample hashes and static / dynamic analysis results
  • Model verdicts with prompt provenance
  • Independent analyst and deterministic scanner verdicts

Coverage limit. This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.

Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.

Sigma
Draft SigmaDownload Sigma
title: Hunt - Independent malware-analysis validation
id: aa44bd2e-f46c-4884-a6db-e6b370a12c50
status: experimental
description: Instructions inside a sample influence an LLM analysis result, or
  obfuscation changes a model's classification without changing malicious
  behavior.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
  - Authorized activity matching this broad hunt; baseline and investigate
    before alerting.
level: low
logsource:
  category: application
  product: openattacks_audit
  definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
    Target and CorrelationId from the required source logs first.
detection:
  selection:
    EventType:
      - analysis.verdict
  condition: selection
KQL
Draft KQLDownload KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("analysis.verdict")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated desc
SPL
Draft SPLDownload SPL
sourcetype="openattacks:audit"
EventType IN ("analysis.verdict")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _time
Investigation steps
  1. Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
  2. Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
03MATCHBOIL still installsUnexpected payload / remote-access behaviorCustom schema · Draft

If the script is run, it still downloads MATCHBOIL, the group's loader. CERT-UA's July chain uses a double-extension VBS, Notepad++ components, LUNCHPOKE, BURNYBEAR, and MATCHBOIL.V2.

What to look for

A program delivered by the reported lure executes, establishes access or contacts infrastructure outside the approved software baseline.

If the script runs, it still downloads MATCHBOIL. The comment does not change that behavior.

Visibility: Defender-side logs where collected

Required logs

  • Platform-specific process and file provenance
  • Signer / hash and software inventory
  • Per-process network and remote-session records

Coverage limit. A program name alone is not a signature. This normalized template requires the relevant platform collector and does not assert Windows-only policy coverage.

Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.

Sigma
Draft SigmaDownload Sigma
title: Hunt - Unexpected payload / remote-access behavior
id: 998d4263-6e73-4472-a74d-cb4d61c28544
status: experimental
description: A program delivered by the reported lure executes, establishes
  access or contacts infrastructure outside the approved software baseline.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
  - Authorized activity matching this broad hunt; baseline and investigate
    before alerting.
level: low
logsource:
  category: application
  product: openattacks_audit
  definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
    Target and CorrelationId from the required source logs first.
detection:
  selection:
    EventType:
      - process.unapproved.exec
      - remote.session.start
  condition: selection
KQL
Draft KQLDownload KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("process.unapproved.exec", "remote.session.start")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated desc
SPL
Draft SPLDownload SPL
sourcetype="openattacks:audit"
EventType IN ("process.unapproved.exec", "remote.session.start")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _time
Investigation steps
  1. Confirm the payload's operating system, origin and execution context before choosing a platform-specific rule.
  2. Correlate the launched program with child processes, persistence and network activity; verify authorization.
04A refusal is the gapIndependent malware-analysis validationCustom schema · Draft

ESET did not measure the effect. DugganUSA's separate one-day test saw one of three models return an empty refusal and the other two flag the decoy and continue.

What to look for

Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.

DugganUSA's one-day test is independent of ESET and used a synthetic script. One refusal is not a measurement of this incident file.

Visibility: Review / research workflow

Required logs

  • Sample hashes and static / dynamic analysis results
  • Model verdicts with prompt provenance
  • Independent analyst and deterministic scanner verdicts

Coverage limit. This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.

Custom-schema candidates, not out-of-box detectors. Map your application / cloud / platform logs to EventType, Actor, Asset, Action, Target and CorrelationId. KQL uses a custom OpenAttacksAudit_CL table; SPL uses openattacks:audit. EventType values are proposed normalization names, not native vendor event names. Correlation and environment-specific alert logic remain to be built.

Sigma
Draft SigmaDownload Sigma
title: Hunt - Independent malware-analysis validation
id: aa44bd2e-f46c-4884-a6db-e6b370a12c50
status: experimental
description: Instructions inside a sample influence an LLM analysis result, or
  obfuscation changes a model's classification without changing malicious
  behavior.
author: MagicSword OpenAttacks catalog
date: 2026-10-23
falsepositives:
  - Authorized activity matching this broad hunt; baseline and investigate
    before alerting.
level: low
logsource:
  category: application
  product: openattacks_audit
  definition: Custom normalized audit schema; map EventType, Actor, Asset, Action,
    Target and CorrelationId from the required source logs first.
detection:
  selection:
    EventType:
      - analysis.verdict
  condition: selection
KQL
Draft KQLDownload KQL
// SCHEMA TEMPLATE: create/map this custom table before use; not a built-in Sentinel table.
// OpenAttacksAudit_CL: TimeGenerated(datetime), EventType/Actor/Asset/Action/Target/CorrelationId(string).
// Selects candidates only. Apply the stage hypothesis and correlation steps before alerting.
OpenAttacksAudit_CL
| where TimeGenerated > ago(7d)
| where EventType in ("analysis.verdict")
| project TimeGenerated, EventType, Actor, Asset, Action, Target, CorrelationId
| order by TimeGenerated desc
SPL
Draft SPLDownload SPL
sourcetype="openattacks:audit"
EventType IN ("analysis.verdict")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _time
Investigation steps
  1. Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
  2. Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
Technical detailsTooling, impact & MITRE ATLAS

Attack context

AI's role
Other AI-enabled attacks
Initial access
A VBScript used in the early stage of an intrusion at a Ukrainian organization. ESET does not describe how the script arrived. CERT-UA's July chain for the same group starts from a phishing ZIP that contains a VBS file disguised with a fake extension.

Tooling as reported

AI products, models and artifacts

  • Unnamed LLM-powered code scanner. ESET did not name a model.
  • DugganUSA's independent test used Claude Sonnet 4.5, GPT-4o, and Mistral. That test is not the incident.

Other tooling

  • MATCHBOIL loader, delivered by the script
  • Windows Script Host (wscript.exe or cscript.exe)

Targets and reported impact

Targets / sectors
An unnamed organization in Ukraine. UAC-0099's usual targets include transportation and energy.
Reported scale / exposure
One documented script in one early-stage intrusion. ESET did not report how often the comment has been used or whether a scanner refused this sample.
Target geography
Ukraine

Ukraine is the victim location ESET gives for this script. Russia is the actor alignment, not a victim region.

Reported attribution

UAC-0099, Russia-aligned, espionage. ESET attributes the script and the GuardBreaker comment. CERT-UA attributes the group's MATCHBOIL, LUNCHPOKE, and BURNYBEAR tooling. ESET does not claim this sample was handed to Sandworm.

Structured attribution

Actor
UAC-0099
Country
Russia
Motivation
nation-state
Attributed by
ESET Research; CERT-UA for the group and its loaders
Attribution confidence
high

Detection rules

  • Cscript/Wscript Uncommon Script Extension ExecutionSIGMAView rule

    Selected from a security-detections-mcp search for wscript. Windows process creation. wscript.exe or cscript.exe whose command line contains a non-script extension such as .pdf, .doc, or .txt. CERT-UA described UAC-0099 VBS names with a fake document extension, spaces, and then .vbs. A command line that still contains .pdf matches. A plain .vbs path with no fake extension does not.

  • Suspicious Double Extension FilesSIGMAView rule

    Selected from a security-detections-mcp search for double extension. Windows file event. A created file whose name ends in a document extension immediately followed by .vbs, .exe, or another listed payload extension, including .pdf.vbs. The spaced form described by CERT-UA, with characters between pdf and .vbs, does not match this regex.

  • Vbscript Execution Using Wscript AppSPLUNKView rule

    Selected from a security-detections-mcp search for VBScript. Endpoint process data for wscript.exe executing VBScript. This is the generic script-host path for the GuardBreaker file. It does not know that the comment is a refusal bait.

  • Refusal-bait comments and MATCHBOILGUIDANCE

    security-detections-mcp searches for VBScript, wscript, and double extension did not return a rule that looks for refusal-bait wording or for MATCHBOIL. A YARA or Sigma hunt for inert comments that request help building a weapon, sitting above a downloader, is still a written hunt rather than a published rule. Do not treat a scanner refusal as a benign verdict. CERT-UA's July chain also drops Notepad++ components and LUNCHPOKE, BURNYBEAR, and MATCHBOIL.V2; those names are not in the indexed rules.

Activity timeline

First seen
2026-08-27
Last updated
2026-09-10

Updates

  • 2026-07-21 — CERT-UA

    CERT-UA documents UAC-0099's updated chain, including a double-extension VBS, Notepad++ 8.8.3 components, LUNCHPOKE, BURNYBEAR, and MATCHBOIL.V2. The advisory does not describe the refusal-bait comment.

  • 2026-08-27 — ESET Research

    ESET posts the GuardBreaker finding on X: a UAC-0099 VBScript used early against a target in Ukraine carries an inert comment meant to make an LLM code scanner refuse.

  • 2026-08-31 — Help Net Security

    Help Net Security recounts the X posts, the nuclear-weapon comment, and the MATCHBOIL downloader.

  • 2026-09-01 — DugganUSA

    Independent A/B test on a synthetic script. One of three models returned an empty refusal. The other two analyzed the sample and flagged the decoy. DugganUSA says this is not a benchmark.

  • 2026-09-10 — WeLiveSecurity

    Tomáš Foltýn publishes the GuardBreaker article. ESET still does not report a test of the comment or name a targeted model.

Source defensive notes

Keep the scanner's instructions separate from the file under analysis, and treat text inside the sample as data. A refusal, an empty completion, or a safety block on a malware sample should raise a review, not end it. Fall back to analysis that does not depend on a model, and do not give one model the only vote on whether a file is safe. The comment does not change what the script does when Windows runs it, so script-host and loader detections still apply.

MITRE ATLAS

AI-specific catalog annotations based on the reported behavior; these are not MITRE case-study assignments.

AML.T0051.001Indirect

The refusal bait is indirect prompt injection: untrusted text inside a file being analyzed is meant to reach the model and cut the analysis short. Source

AML.T0015Evade AI Model

The intended effect is evasion of an AI model used as a code scanner, by provoking a safety refusal before the payload logic is read. Source

Sources & research credit

4 sources

Research by ESET Research · ESET

  1. GuardBreaker: derailing AI-assisted malware analysis with a code comment www.welivesecurity.com
  2. Russian hackers plant a nuclear-weapon prompt in malware to trip AI safety filters www.helpnetsecurity.com
  3. UAC-0099: LUNCHPOKE, BURNYBEAR, updated MATCHBOIL.V2, and Notepad++ 8.8.3 cert.gov.ua
  4. Independent A/B test of the decoy comment on three models www.dugganusa.com

The comment's role, the MATCHBOIL purpose, the Ukraine victim, and the IDA and Wireshark checks are from the WeLiveSecurity article (Sep 10, 2026) and the Aug 31 Help Net Security recap of the X posts. The enrichment procedure is not repeated here. DugganUSA is independent research on a synthetic sample; it is not a test of the incident file. CERT-UA article 6318634 (Jul 21, 2026) is the government write-up of the group's VBS, Notepad++, LUNCHPOKE, BURNYBEAR, and MATCHBOIL.V2 chain. https://cybernews.com/security/russian-hackers-nuclear-prompt-ai-guardrails/ returned HTTP 403 during link checks and is omitted. No YARA or Sigma rule in the detection index matches the refusal-bait wording or the name MATCHBOIL.

Record & downloads
Source record name
ESET GuardBreaker - UAC-0099 plants a refusal-bait comment in a MATCHBOIL script
Entry ID
eset-guardbreaker-uac-0099-decoy-prompt-2026-08
Catalog review
November 17, 2026
Category
attacks-on-ai
AI involvement confidence
vendor-asserted