Bringing Transparency and Accountability to
AI System Prompts

Dimension Coverage Across Products

Percentage of audited prompts that contain at least one annotation per dimension

Across 88 audited system prompts, the share carrying at least one instruction on each AISPA dimension — Identity Transparency 81.8% protective and 3.4% problematic; Truthfulness & Information Integrity 94.3% protective and 14.8% problematic; Privacy & Data Protection 62.5% protective and 2.3% problematic; Tool/Action Safety 72.7% protective and 10.2% problematic; User Agency & Manipulation Prevention 92.0% protective and 18.2% problematic; Unsafe Request Handling 60.2% protective and 6.8% problematic; Harm Prevention & User Safety 67.0% protective and 10.2% problematic; Fairness, Inclusion & Neutrality 62.5% protective and 4.5% problematic.

User Protection Evolution
Across Representative Series

Number of protective and problematic instruction entries per product version, grouped by organization

Anthropic, version by version: Claude-3.5-Sonnet-2024-07-12 (2024-07) 26 protective, 0 problematic; Claude-3.7-Sonnet-2025-02-25 (2025-02) 36 protective, 1 problematic; Claude-Sonnet-4-2025-06-03 (2025-06) 56 protective, 0 problematic; Claude-Sonnet-4.5-2025-10-28 (2025-10) 77 protective, 0 problematic; Claude-Opus-4.6-2025-05-14 (2025-05) 81 protective, 0 problematic; Claude Sonnet 4.6 (None) 85 protective, 1 problematic; Claude Opus 4.7 (None) 102 protective, 0 problematic; Claude Opus 4.8 (None) 125 protective, 0 problematic; Claude Sonnet 5 (None) 128 protective, 2 problematic; Claude Opus 5 (None) 149 protective, 1 problematic; Claude Fable 5 (None) 159 protective, 0 problematic.

OpenAI, version by version: GPT-4o-2025-07-29 (2025-07) 25 protective, 0 problematic; o3-2025-06-04 (2025-06) 29 protective, 1 problematic; GPT-5-2025-08-07 (2025-08) 35 protective, 0 problematic; GPT-5-Thinking-2025-08-23 (2025-08) 66 protective, 0 problematic; GPT-5.2-Thinking-2025-12-13 (2025-12) 83 protective, 0 problematic; GPT-5.3 (None) 48 protective, 0 problematic; GPT-5.4 Thinking (None) 63 protective, 0 problematic; GPT-5.5 Thinking (None) 135 protective, 6 problematic; GPT-5.6 Sol (extra-high) (None) 105 protective, 5 problematic.

xAI, version by version: Grok-1-2023-12-14 (2023-12) 5 protective, 4 problematic; Grok-2-2024-08-21 (2024-08) 8 protective, 6 problematic; Grok-3-2025-05-16 (2025-05) 11 protective, 2 problematic; Grok-4-2025-10-27 (2025-10) 9 protective, 0 problematic; Grok 4.1 (None) 19 protective, 2 problematic; Grok-4.2-2026-02-17 (2026-02) 21 protective, 0 problematic; Grok 4.3 (None) 34 protective, 2 problematic; Grok 4.5 (None) 39 protective, 2 problematic.