A3 malware that tells AI analysts to look away (Talos)
84 malware samples embed comments telling LLM analysts to ignore or refuse the file. In the best runs, the trick steered models toward a benign verdict about 35% of the time.
What happened
84 samples across families embed plaintext comments telling LLMs to ignore the file, refuse reverse engineering, or claim DoD/IP. Direct-instruction trick steered models toward benign ~35% in best runs. MANTLEMAZE pairs this with iqvw64e.sys BYOVD (CVE-2015-2291); ROZESHELL pairs with AMSI bypass and csc.exe.
Malware embeds natural-language instructions to steer defender LLM triage (prompt injection against analyzers)
How it happened
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.
- 01 · Malware is presented for analysis
Samples include comments aimed at automated analysis models.
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 1 - 02 · Instructions target the analyst model
The text asks LLMs to ignore the file, refuse analysis or accept benign explanations.
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 - 03 · Model interpretation is influenced
The research measures whether direct instructions shift maliciousness assessments.
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 3 - 04 · Some runs misclassify samples
Results are bounded by the evaluated models, samples and experimental conditions.
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: unknown (varies by family; technique is anti-analysis layer). Reported impact: Results are bounded by the evaluated models, samples and experimental conditions.
Check these source-linked conditions against your setup.
The malware samples contain instructions aimed at LLM-based analysis pipelines.
Conditions to check- Your workflow must submit attacker-controlled sample content to an LLM analyst.
- The reported model-evasion measurements do not establish that your analysis tool will behave identically.
Source-linked applicability guidance. Confirm current scope in the original report.
What you can do
Protection & prevention
Treat embedded sample text as evidence, isolate it from analyst instructions and validate model verdicts independently.
Treat sample text as evidence never instruction; prevent iqvw64e load via app control
Catalog defensive guidance based on the reported behavior. Confirm scope and current mitigations in the original source.
Detection queries
Draft query starters grouped by the attack stages where they apply.
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.
01Malware is presented for analysisIndependent malware-analysis validationCustom schema · Draft
Samples include comments aimed at automated analysis models.
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 separate MANTLEMAZE driver and ROZESHELL compiler behaviors are not edges in this analysis-evasion flow; do not infer a BYOVD chain from this node.
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
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
// 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 descSPL
sourcetype="openattacks:audit"
EventType IN ("analysis.verdict")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
02Instructions target the analyst modelIndependent malware-analysis validationCustom schema · Draft
The text asks LLMs to ignore the file, refuse analysis or accept benign explanations.
What to look for
Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.
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
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
// 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 descSPL
sourcetype="openattacks:audit"
EventType IN ("analysis.verdict")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
03Model interpretation is influencedIndependent malware-analysis validationCustom schema · Draft
The research measures whether direct instructions shift maliciousness assessments.
What to look for
Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.
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
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
// 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 descSPL
sourcetype="openattacks:audit"
EventType IN ("analysis.verdict")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
04Some runs misclassify samplesIndependent malware-analysis validationCustom schema · Draft
Results are bounded by the evaluated models, samples and experimental conditions.
What to look for
Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.
Evaluate disagreements against independently labeled samples and benign controls; results apply to the tested models and conditions.
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
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
// 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 descSPL
sourcetype="openattacks:audit"
EventType IN ("analysis.verdict")
| table _time EventType Actor Asset Action Target CorrelationId
| sort - _timeInvestigation steps
- Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- 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
- unknown (varies by family; technique is anti-analysis layer)
Tooling as reported
AI products, models and artifacts
- Prompt injection against defender LLM pipelines
Other tooling
- iqvw64e.sys (LOLDrivers)
- csc.exe (LOLBAS)
- VMProtect costume sections
Targets and reported impact
- Reported scale / exposure
- 84 samples Jan 2025–Jul 2026; ~35% steering in best direct-instruction runs
Reported attribution
CAIRN families FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE, ROZESHELL; A3 technique class
Structured attribution
- Actor
- CAIRN families FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE, ROZESHELL
Detection rules
- Review attack-specific indicatorsGUIDANCE
Monitor for the techniques described in A3 AI-analysis evasion (FRUITSHELL / HOLLOWCLAD / MANTLEMAZE). Refer to sources for specific IOCs and TTPs.
Source defensive notes
Treat sample text as evidence never instruction; prevent iqvw64e load via app control
MITRE ATLAS
AI-specific catalog annotations based on the reported behavior; these are not MITRE case-study assignments.
Comments in analyzed files inject instructions into a security-analysis model. Source
The text attempts to evade AI-based maliciousness assessment. Source
Sources & research credit
1 sourceResearch by Ryan Fetterman · Cisco Talos (CAIRN)
Record & downloads
- Source record name
- A3 AI-analysis evasion (FRUITSHELL / HOLLOWCLAD / MANTLEMAZE)
- Entry ID
- talos-a3-ai-analysis-evasion-2026-10
- Catalog review
- October 20, 2026
- Category
- adversary-use-of-ai
- AI involvement confidence
- confirmed