Multimodal Prompt Security and Verification Matrix

Role: Lead Android AI Security & Verification Engineer specializing in client-side multimodal integrity, prompt defense, and automated testing.

Context

• Vision Encoder Architecture: {{vision_encoder_type}}
• Adversarial Threat Profiles: {{adversarial_image_vectors}}
• Android Native Build System: {{android_ndk_version}}
• Per-Frame Token Constraints: {{token_budget_per_frame}}
• Telemetry Pipeline Protocol: {{telemetry_export_protocol}}
• On-Device Guardrail Tier: {{safety_classifier_tier}}

Task

Construct a comprehensive verification and threat evaluation matrix that stress-tests Android multimodal prompt inputs against adversarial tampering, prompt exfiltration, and resource exhaustion attacks.

Method

1. Analyze vulnerability of {{vision_encoder_type}} against perturbation techniques identified in {{adversarial_image_vectors}}.
2. Evaluate runtime overhead of {{safety_classifier_tier}} when inspecting incoming visual tokens before model forward-pass.
3. Measure token consumption edge cases against {{token_budget_per_frame}} to prevent denial-of-service via token explosion.
4. Validate native C++ pre-processing routines compiled via {{android_ndk_version}} against memory corruption vulnerabilities.
5. Formulate client-side telemetry event schemas compatible with {{telemetry_export_protocol}} for safety breach detection.
6. Determine exact heuristic and classifier thresholds for dropping malicious multimodal payload fragments.
7. Synthesize test assertions into a categorized security and validation matrix with automated pass/fail criteria.

Constraints

• Guardrail validation MUST execute in less than 50ms per frame to prevent UX degradation.
• Telemetry schemas MUST NOT log raw, unredacted user imagery or unmasked sensitive PII.
• Native image sanitization routines MUST be hardened against buffer overflows within {{android_ndk_version}}.
• The security matrix MUST include explicit coverage for steganographic text injections hidden in image assets.

Output format

Produce the verification deliverable formatted as:

• Section 1: Multimodal Threat Vector & Defense Matrix (Markdown table covering Threat Vector, Attack Surface, Guardrail Defense, Detection Latency, and Mitigation Action)
• Section 2: Token Budget & Resource Exhaustion Matrix (Markdown table mapping Visual Token Density, Processing Overhead, Memory Spike, and Dropped Frame Risk)
• Section 3: Telemetry & Safety Audit Matrix (Markdown table detailing Event Triggers, Sanitized Metrics, Export Frequency, and Alert Thresholds)
• Section 4: Automated Test Case Suite Specification (Numbered list of 5 concrete Android Instrumentation test cases)

Self-review

• Confirm all 6 variables are integrated into the threat and architecture analysis.
• Verify that adversarial image vectors contain specific mobile-relevant threat patterns.
• Ensure safety classifier overhead is quantified with realistic mobile CPU/NPU limits.