| """Cases taken from each model's own training corpus, loaded from cases.json. |
| |
| Every case is an attack, and every one is a real record rather than something |
| written for the demo: the anomaly sessions come from the agent_trust split of |
| `AnShi-HW/tool-call-anomaly-data`, the intent records from tool-intent-align's |
| training corpora, and the output-filter rows from minspan's training split. |
| Each case keeps its source and record id in `cases.json`, and the button label |
| is the attack class that corpus assigns it, so the labels stay short and read |
| as one set. |
| |
| `scripts/extract_cases.py` regenerates the file. |
| """ |
|
|
| from __future__ import annotations |
|
|
| import json |
| from pathlib import Path |
|
|
| CASES = json.loads(Path(__file__).with_name("cases.json").read_text(encoding="utf-8")) |
|
|
|
|
| def pretty(value): |
| return json.dumps(value, ensure_ascii=False, indent=2) |
|
|
|
|
| def labels(task): |
| return [(case["label"], "risk") for case in CASES[task]] |
|
|
|
|
| INTENT_CASES = CASES["intent"] |
| OUTPUT_CASES = CASES["output_filter"] |
| ANOMALY_CASES = CASES["anomaly"] |
|
|
| INTENT_LABELS = labels("intent") |
| OUTPUT_LABELS = labels("output_filter") |
| ANOMALY_LABELS = labels("anomaly") |
|
|
| INTENT_EXAMPLES = [ |
| [pretty(case["context"]), pretty(case["tool"]), pretty(case["call"])] for case in INTENT_CASES |
| ] |
| OUTPUT_EXAMPLES = [[case["instruction"], case["data"]] for case in OUTPUT_CASES] |
| ANOMALY_EXAMPLES = [[case["task"], pretty(case["tool_calls"])] for case in ANOMALY_CASES] |
|
|
