# A3 AI-analysis evasion (FRUITSHELL / HOLLOWCLAD / MANTLEMAZE)

- ID: talos-a3-ai-analysis-evasion-2026-10
- Reported: 2026-10-08
- Researcher: Ryan Fetterman
- Organization: Cisco Talos (CAIRN)
- AI role: other (Other AI-enabled attacks)
- Confidence in AI involvement: not assessed
- Research / borderline: false
- Last reviewed: 2026-10-20
- Entry page: https://openattacks.ai/attacks/talos-a3-ai-analysis-evasion-2026-10/
- Entry JSON: https://openattacks.ai/api/attacks/talos-a3-ai-analysis-evasion-2026-10.json
- Browse title: A3 malware that tells AI analysts to look away (Talos)
- Editorial summary: 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.
- Evidence class (catalog annotation): Malware analysis
- Source-stated target environments: Not specified
- Stage hunts and conditional defenses: https://openattacks.ai/api/defenses/talos-a3-ai-analysis-evasion-2026-10.json

## Technique
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.

## Stack relevance (editorial)
Assessment: mapped. Editorial applicability reviewed against the catalog source account; not a scan or independent incident verification.
First action: Treat embedded sample text as evidence, isolate it from analyst instructions and validate model verdicts independently.
- Relationship: workflow; tools: ai-analysis; OS scope: not OS-specific / not stated; requires existing MCP connections: false.
  Relevance: The malware samples contain instructions aimed at LLM-based analysis pipelines.
  Condition to check: Your workflow must submit attacker-controlled sample content to an LLM analyst.
  Condition to check: The reported model-evasion measurements do not establish that your analysis tool will behave identically.
  Evidence: https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/
A product mention alone is not a victim-side applicability mapping. Selection does not verify affected versions, permissions or compromise.

## Initial access
unknown (varies by family; technique is anti-analysis layer)

## AI tooling and models
- Prompt injection against defender LLM pipelines

## Other tooling
- iqvw64e.sys (LOLDrivers)
- csc.exe (LOLBAS)
- VMProtect costume sections

## Victims and sectors
unknown

## Target geography
- unknown

## Attribution
CAIRN families FRUITSHELL, PLOTSAFE, HOLLOWCLAD, MANTLEMAZE, ROZESHELL; A3 technique class

## Scale and metrics
84 samples Jan 2025–Jul 2026; ~35% steering in best direct-instruction runs

## Attack flow
Catalog reconstruction of the supplied source account. Unknowns and research caveats are preserved.
### step-1: Malware is presented for analysis [access]
Samples include comments aimed at automated analysis models.
Evidence: https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/

### step-2: Instructions target the analyst model [ai]
The text asks LLMs to ignore the file, refuse analysis or accept benign explanations.
Evidence: https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/

### step-3: Model interpretation is influenced [execution]
The research measures whether direct instructions shift maliciousness assessments.
Evidence: https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/

### step-4: Some runs misclassify samples [impact]
Results are bounded by the evaluated models, samples and experimental conditions.
Evidence: https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/

Connections:
- step-1 -> step-2
- step-2 -> step-3
- step-3 -> step-4

## Stage defense and hunting recommendations
Editorial recommendations, not validated coverage. Source-controlled query starters are drafts. Custom-schema examples require telemetry collection and field mapping; absence of a mapping means not assessed. MagicSword policy opportunities require managed Windows scope and enforcement.
### step-1: Independent malware-analysis validation
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/analysis-review.sigma.yml
- KQL query download: /api/detections/analysis-review.kql
- SPL query download: /api/detections/analysis-review.spl
- Hunt hypothesis: Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.
- Required telemetry: Sample hashes and static / dynamic analysis results; Model verdicts with prompt provenance; Independent analyst and deterministic scanner verdicts
- Investigate: Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- Investigate: Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
- Limitation: This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.
- Qualification: 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.
- No direct MagicSword prevention claim is recorded for this stage.
- Community intelligence: LOLDrivers — https://www.loldrivers.io

### step-2: Independent malware-analysis validation
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/analysis-review.sigma.yml
- KQL query download: /api/detections/analysis-review.kql
- SPL query download: /api/detections/analysis-review.spl
- Hunt hypothesis: Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.
- Required telemetry: Sample hashes and static / dynamic analysis results; Model verdicts with prompt provenance; Independent analyst and deterministic scanner verdicts
- Investigate: Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- Investigate: Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
- Limitation: This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.
- No direct MagicSword prevention claim is recorded for this stage.

### step-3: Independent malware-analysis validation
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/analysis-review.sigma.yml
- KQL query download: /api/detections/analysis-review.kql
- SPL query download: /api/detections/analysis-review.spl
- Hunt hypothesis: Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.
- Required telemetry: Sample hashes and static / dynamic analysis results; Model verdicts with prompt provenance; Independent analyst and deterministic scanner verdicts
- Investigate: Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- Investigate: Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
- Limitation: This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.
- No direct MagicSword prevention claim is recorded for this stage.

### step-4: Independent malware-analysis validation
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/analysis-review.sigma.yml
- KQL query download: /api/detections/analysis-review.kql
- SPL query download: /api/detections/analysis-review.spl
- Hunt hypothesis: Instructions inside a sample influence an LLM analysis result, or obfuscation changes a model's classification without changing malicious behavior.
- Required telemetry: Sample hashes and static / dynamic analysis results; Model verdicts with prompt provenance; Independent analyst and deterministic scanner verdicts
- Investigate: Compare model conclusions with independent behavioral evidence and a controlled evaluation set.
- Investigate: Treat sample comments as data; measure disagreement and false negatives without following embedded instructions.
- Limitation: This evaluates analysis robustness, not victim endpoint compromise. Side behaviors mentioned elsewhere in a report do not establish edges in this flow.
- Qualification: Evaluate disagreements against independently labeled samples and benign controls; results apply to the tested models and conditions.
- No direct MagicSword prevention claim is recorded for this stage.


## MITRE ATLAS
Mappings are catalog annotations based on described behavior, not assignments or endorsements by MITRE.
- AML.T0051.001: Indirect — Comments in analyzed files inject instructions into a security-analysis model.
  Technique: https://atlas.mitre.org/techniques/AML.T0051.001
  Evidence: https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/
- AML.T0015: Evade AI Model — The text attempts to evade AI-based maliciousness assessment.
  Technique: https://atlas.mitre.org/techniques/AML.T0015
  Evidence: https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/

## Defensive notes
Treat sample text as evidence never instruction; prevent iqvw64e load via app control

## Sources
- https://blog.talosintelligence.com/ignore-all-instructions-and-read-this-blog-the-state-of-ai-analysis-evasion-in-malware/
