# Pretext LLM attacker vs agent-skill scanners

- ID: pretext-skill-scanner-evasion-2026-10
- Reported: 2026-10-05
- Researcher: Tobias Kaisar, Aritra Dhar
- Organization: Huawei Research Zurich
- AI role: other (Other AI-enabled attacks)
- Confidence in AI involvement: not assessed
- Research / borderline: true
- Last reviewed: 2026-10-20
- Entry page: https://openattacks.ai/attacks/pretext-skill-scanner-evasion-2026-10/
- Entry JSON: https://openattacks.ai/api/attacks/pretext-skill-scanner-evasion-2026-10.json
- Browse title: Pretext: LLM attacker evades agent-skill scanners
- Editorial summary: Academic research: plain-language payloads split across skill files evaded skill scanners up to 97% of the time, showing whole bundles must be scanned. This is not a real intrusion.
- Evidence class (catalog annotation): Research demonstration
- Source-stated target environments: Agents & MCP
- Stage hunts and conditional defenses: https://openattacks.ai/api/defenses/pretext-skill-scanner-evasion-2026-10.json

## Evidence qualification
Academic research paper on evading skill scanners, not a real-world intrusion.

## Technique
LLM attacker wrote payloads as plain-language instructions, framed as skill purpose, split across files. Evaded NVIDIA SkillSpector-style static+LLM-judge scanners up to 97% vs fixed detector, 77% vs adaptive. Takeaway: scanning SKILL.md alone is insufficient; scan whole bundle.

## Stack relevance (editorial)
Assessment: mapped. Editorial applicability reviewed against the catalog source account; not a scan or independent incident verification.
First action: Review the complete skill bundle and isolate skill permissions; inspect instructions spread across files.
- Relationship: workflow; tools: agent-skills; OS scope: not OS-specific / not stated; requires existing MCP connections: false.
  Relevance: The academic research splits instructions across skill files to evade scanners.
  Condition to check: An agent must load the skill bundle and process its instructions.
  Condition to check: The reported evasion results come from academic research rather than a confirmed intrusion.
  Evidence: https://arxiv.org/abs/2609.39607
A product mention alone is not a victim-side applicability mapping. Selection does not verify affected versions, permissions or compromise.

## Initial access
n/a (research)

## AI tooling and models
- LLM attacker
- agent skills / SKILL.md
- [NVIDIA SkillSpector](https://github.com/NVIDIA/SkillSpector) (evaluated)

## Other tooling
- unknown

## Victims and sectors
n/a (research against scanners)

## Target geography
- Not applicable
Controlled research against skill scanners, rather than a geographically located victim campaign.

## Attribution
Academic (Pretext paper)

## Scale and metrics
Up to 97% / 77% evasion rates in paper evaluation

## Attack flow
Catalog reconstruction of the supplied source account. Unknowns and research caveats are preserved.
### step-1: Researchers construct skill bundles [access]
The experiment evaluates skill-scanner resistance rather than a victim intrusion.
Evidence: https://arxiv.org/abs/2609.39607

### step-2: LLM writes pretextual instructions [ai]
Payloads are framed as ordinary skill purpose and split across files.
Evidence: https://arxiv.org/abs/2609.39607

### step-3: Static and model-based scanners inspect them [execution]
Tests compare evasion against fixed and adaptive detectors.
Evidence: https://arxiv.org/abs/2609.39607

### step-4: Research measures scanner evasion [impact]
Reported evasion rates apply to the evaluated detectors and experimental settings.
Evidence: https://arxiv.org/abs/2609.39607

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: Skill bundle provenance / review
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/skill-review.sigma.yml
- KQL query download: /api/detections/skill-review.kql
- SPL query download: /api/detections/skill-review.spl
- Hunt hypothesis: A skill bundle introduces new network, execution or secret-access instructions, including instructions distributed across non-Markdown files.
- Required telemetry: Complete skill-bundle inventory and hashes; Scanner results, extracted attachments and review history; Agent load / tool logs if execution is actually observed
- Investigate: Review the entire bundle, including images and referenced files, in an isolated static-analysis workflow.
- Investigate: Compare requested capabilities with the skill's purpose; record scanner disagreement and avoid treating a finding as proof of execution.
- Limitation: Scanner findings and evasion results are research evidence, not a runtime compromise detector. Do not execute embedded commands or invent a victim chain.
- No direct MagicSword prevention claim is recorded for this stage.
- Community intelligence: LOLSkills — https://lolskills.io

### step-2: Skill bundle provenance / review
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/skill-review.sigma.yml
- KQL query download: /api/detections/skill-review.kql
- SPL query download: /api/detections/skill-review.spl
- Hunt hypothesis: A skill bundle introduces new network, execution or secret-access instructions, including instructions distributed across non-Markdown files.
- Required telemetry: Complete skill-bundle inventory and hashes; Scanner results, extracted attachments and review history; Agent load / tool logs if execution is actually observed
- Investigate: Review the entire bundle, including images and referenced files, in an isolated static-analysis workflow.
- Investigate: Compare requested capabilities with the skill's purpose; record scanner disagreement and avoid treating a finding as proof of execution.
- Limitation: Scanner findings and evasion results are research evidence, not a runtime compromise detector. Do not execute embedded commands or invent a victim chain.
- No direct MagicSword prevention claim is recorded for this stage.

### step-3: Skill bundle provenance / review
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/skill-review.sigma.yml
- KQL query download: /api/detections/skill-review.kql
- SPL query download: /api/detections/skill-review.spl
- Hunt hypothesis: A skill bundle introduces new network, execution or secret-access instructions, including instructions distributed across non-Markdown files.
- Required telemetry: Complete skill-bundle inventory and hashes; Scanner results, extracted attachments and review history; Agent load / tool logs if execution is actually observed
- Investigate: Review the entire bundle, including images and referenced files, in an isolated static-analysis workflow.
- Investigate: Compare requested capabilities with the skill's purpose; record scanner disagreement and avoid treating a finding as proof of execution.
- Limitation: Scanner findings and evasion results are research evidence, not a runtime compromise detector. Do not execute embedded commands or invent a victim chain.
- No direct MagicSword prevention claim is recorded for this stage.

### step-4: Skill bundle provenance / review
- Visibility: Review / research workflow
- Query origin: catalog_authored; status: draft; log scope: custom_schema
- SIGMA query download: /api/detections/skill-review.sigma.yml
- KQL query download: /api/detections/skill-review.kql
- SPL query download: /api/detections/skill-review.spl
- Hunt hypothesis: A skill bundle introduces new network, execution or secret-access instructions, including instructions distributed across non-Markdown files.
- Required telemetry: Complete skill-bundle inventory and hashes; Scanner results, extracted attachments and review history; Agent load / tool logs if execution is actually observed
- Investigate: Review the entire bundle, including images and referenced files, in an isolated static-analysis workflow.
- Investigate: Compare requested capabilities with the skill's purpose; record scanner disagreement and avoid treating a finding as proof of execution.
- Limitation: Scanner findings and evasion results are research evidence, not a runtime compromise detector. Do not execute embedded commands or invent a victim chain.
- Qualification: Compare scanner results and independent review on the evaluated bundles; measured evasion is not a runtime compromise alert.
- 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 — Skill-bundle text carries instructions targeting the model-based scanner.
  Technique: https://atlas.mitre.org/techniques/AML.T0051.001
  Evidence: https://arxiv.org/abs/2609.39607
- AML.T0015: Evade AI Model — The research measures evasion of AI-based detectors.
  Technique: https://atlas.mitre.org/techniques/AML.T0015
  Evidence: https://arxiv.org/abs/2609.39607

## Defensive notes
Whole-bundle skill scanning (LOLSkills model)

## Sources
- https://arxiv.org/abs/2609.39607

## Published MagicSword coverage
- https://x.com/_josehelps/status/2107213992989217128
- https://x.com/_josehelps/status/2107213992989217128
