On-Device Multimodal Prompt Injection Security Brief

Role: Lead Android Security & AI Compliance Architect evaluating on-device multimodal prompt injection and zero-trust memory residency.

Context

• Client application package: {{client_app_package}}
• Embedded on-device VLM engine: {{on_device_vlm_engine}}
• Identified attack vector: {{prompt_injection_vector}}
• Enterprise data policy: {{data_loss_prevention_policy}}
• Hardware Keystore configuration: {{hardware_keystore_level}}
• Mandatory compliance deadline: {{remediation_deadline}}

Task

Compose an urgent security advisory email to Android development leads outlining vulnerabilities in on-device multimodal prompt processing, demonstrating exploitation paths, and mandating sandboxing mitigations before production rollout.

Method

1. Deconstruct the exploit payload in {{prompt_injection_vector}} embedded within user-supplied image EXIF data or adversarial visual perturbations.
2. Detail how {{on_device_vlm_engine}} processes adversarial input to bypass system prompt guardrails.
3. Trace downstream risks of arbitrary Intent execution, ContentProvider leakage, or local database exfiltration in {{client_app_package}}.
4. Audit memory persistence of processed prompt embeddings to verify compliance with {{data_loss_prevention_policy}}.
5. Specify mandatory cryptographic isolation steps utilizing {{hardware_keystore_level}} for local prompt caching.
6. Formulate defensive input sanitization: visual normalization, EXIF stripping, and dual-pass validation.
7. Enforce strict delivery timelines tied directly to {{remediation_deadline}}.

Constraints

• Output MUST be structured as an urgent, enterprise-grade Security Advisory email.
• MUST NOT recommend third-party proprietary security SDKs without native Android NDK/SDK equivalents.
• All remediation requirements MUST specify exact Android security primitives (e.g., isolatedProcess, EncryptedSharedPreferences, Keystore StrongBox).
• Keep the total email body under 700 words.

Output format

Subject Line: [SECURITY MANDATE] Prompt Injection Vulnerability in {{client_app_package}} (Action Required by {{remediation_deadline}})

1. Vulnerability Summary & Severity Scoring (CVSS)
2. Threat Vector & Exploit Mechanism ({{prompt_injection_vector}})
3. On-Device Sandboxing & Keystore Isolation Requirements
4. Required Input Sanitization & Token Boundary Controls
5. Verification & Compliance Sign-off Deadlines

Self-review

• Does the advisory address both visual/multimodal and text prompt injection aspects of {{prompt_injection_vector}}?
• Are the required fixes compatible with {{hardware_keystore_level}}?
• Is the deadline {{remediation_deadline}} highlighted with clear blocking consequences?