| # Source Notes | |
| Use this file to keep short intake notes for each source before and during extraction. | |
| ## Active Sources | |
| - `HackAPrompt` | |
| - `TensorTrust` | |
| ## Current Intake Status | |
| - `HackAPrompt`: `active_normalized` | |
| - `TensorTrust`: `active_normalized` | |
| - `AgentDojo`: `reference_only_examples` | |
| ## Current Active Corpus Summary | |
| - `HackAPrompt` | |
| - 1,000 curated attacks | |
| - included in `data/normalized/attack_cases.jsonl` | |
| - `TensorTrust` | |
| - 1,346 curated attacks | |
| - included in `data/normalized/attack_cases.jsonl` | |
| - Current active normalized corpus total: | |
| - 2,346 attack cases | |
| ## AgentDojo Reference Summary | |
| - Benchmark facts captured from the official paper and repository. | |
| - Source inventory created in `data/curated/agentdojo_source_inventory.json`. | |
| - Source note created in `data/notes/agentdojo_web_extraction.md`. | |
| - AgentDojo examples remain useful for benchmark interpretation, but are not currently part of the active experiment corpus. | |
| ## Note Template | |
| - Source: | |
| - Extraction date: | |
| - Case count imported: | |
| - Attack surfaces covered: | |
| - Ambiguities or label decisions: | |
| - Follow-up needed: | |