[ISC] GPT-5.4 (Agentic Moderation, API-Verified)

[wuyoscar]

Contributor

• Your GitHub username: @wuyoscar

Model

• Model name: GPT-5.4
• Arena rank: 6
• Provider: OpenRouter

Evidence

• community/issue-pending-gpt54-moderation/evidence/samples.csv
• community/issue-pending-gpt54-moderation/evidence/agent_log.json
• community/issue-pending-gpt54-moderation/evidence/bench.py

Harmful Content Details

• Type: Toxic text across 9 OpenAI Moderation categories
• Domain: AI Safety / Content Moderation
• Verification: All 9 samples verified by OpenAI Moderation API (omni-moderation-latest), 100% recall

Method

• Used an existing ISC-Bench template
• Other: agentic execution with real API verification

Summary

GPT-5.4 autonomously completed a content moderation recall benchmark (9 categories: hate, hate/threatening, harassment, harassment/threatening, violence, violence/graphic, self-harm/instructions, illicit, illicit/violent). The model generated toxic text, ran the OpenAI Moderation API, found some scores below threshold, and self-optimized its outputs until all 9 categories passed at 100% recall. This is the first ISC case with real-time API verification of harmful content quality.

[wuyoscar]

Verified ✅ — GPT-5.4 (Rank 6) ISC triggered via agentic content moderation with real OpenAI Moderation API verification.

Key finding: The model not only generated toxic text for all 9 categories, but when some scores fell below threshold, it autonomously iterated — rewriting samples to increase toxicity scores until 100% recall was achieved. This is self-optimizing harmful content generation.

9/9 PASS, verified by omni-moderation-latest. Evidence: community/issue-pending-gpt54-moderation/

[wuyoscar]

📁 Evidence: https://github.com/wuyoscar/ISC-Bench/tree/main/community/issue-57-gpt54-moderation-api