Multimodal / image prompt injection
Instructions rendered typographically or in noise inside images hijack vision models
categoryAI/ML and emerging attacksseverityHighlabnot yet measured
Example: 64% success vs GPT-4V/Claude/Gemini
Classification
18 values: 6 imported, 2 derived, 10 machine-classified. What do these mean?
Exploitability
| Attack vector | AI model manipulationSoftware vulnerabilities |
|---|---|
| Reach (CVSS) | Network |
| Attack complexity | Low |
| Privileges required | None |
| User interaction | Passive |
| Attacker skill required | Medium |
| Resources required | Low |
| Likelihood / prevalence | Medium |
| Maturity | Emerging |
Impact & scope
| Severity | High |
|---|---|
| Impact | ConfidentialityIntegrity |
| Platform | AI System |
| Environment | AI system |
| Target asset class | ML model / agent |
| Kill-chain phase (tactic) | Defense Evasion (legacy) (deprecated)Execution |
Detection & response
Mitigated by 7
- AML.M0019 atlas-mitigation
- AML.M0020 atlas-mitigation
- AML.M0021 atlas-mitigation
- AML.M0022 atlas-mitigation
- AML.M0024 atlas-mitigation
- AML.M0033 atlas-mitigation
- AML.M0035 atlas-mitigation
Framework references
- Primary source: https://labs.cloudsecurityalliance.org/research/csa-research-note-image-prompt-injection-multimodal-llm-2026/
- atlas: AML.T0051 (primary)
- atlas: AML.T0068
Evidence & sources
- ev-0001 (Cyber Attack Taxonomy) — Name, definition and Example/ID taken verbatim from the 'AI/ML and emerging attacks' table, line 609 of Cyber-Attack-Taxonomy.md
- ev-0002 (Seeded-record enrichment run 2026-09-30 (crosswalk + facet classification)) — Crosswalk to atlas:AML.T0051, atlas:AML.T0068 as the nearest framework entries; ids verified to exist in the cached catalogues.