scripts/run_remaining_experiments.sh

#!/usr/bin/env bash
# Resume from experiment #11 — the first 10 already completed on 2026-04-25.
# Runs remaining 28 new experiments + all 14 benign control experiments = 42 total.
# Usage: bash scripts/run_remaining_experiments.sh

set -euo pipefail

PYTHON=".venv312/bin/python"
RUNNER="-m prompt_injection_framework.orchestration.run_experiment"

run() {
    local config="$1"
    echo ""
    echo "════════════════════════════════════════"
    echo "  Running: $config"
    echo "════════════════════════════════════════"
    PYTHONPATH=src $PYTHON $RUNNER --config "$config"
    sleep 45
}

# ── Group 1 (continued): TensorTrust hijacking — groq_compound only ──────────
run configs/tensortrust_hijacking_25_groq_compound.yaml
run configs/tensortrust_hijacking_25_groq_compound_mitigated.yaml

# ── Group 2: HackAPrompt new families bundle (all 7 models) ──────────────────
run configs/hackaprompt_new_families_75.yaml
run configs/hackaprompt_new_families_75_mitigated.yaml

run configs/hackaprompt_new_families_75_llama70b.yaml
run configs/hackaprompt_new_families_75_llama70b_mitigated.yaml

run configs/hackaprompt_new_families_75_qwen32b.yaml
run configs/hackaprompt_new_families_75_qwen32b_mitigated.yaml

run configs/hackaprompt_new_families_75_llama4scout.yaml
run configs/hackaprompt_new_families_75_llama4scout_mitigated.yaml

run configs/hackaprompt_new_families_75_gpt_oss_120b.yaml
run configs/hackaprompt_new_families_75_gpt_oss_120b_mitigated.yaml

run configs/hackaprompt_new_families_75_gpt_oss_20b.yaml
run configs/hackaprompt_new_families_75_gpt_oss_20b_mitigated.yaml

run configs/hackaprompt_new_families_75_groq_compound.yaml
run configs/hackaprompt_new_families_75_groq_compound_mitigated.yaml

# ── Group 3: TensorTrust extraction (3 new models) ───────────────────────────
run configs/tensortrust_extraction_25_gpt_oss_120b.yaml
run configs/tensortrust_extraction_25_gpt_oss_120b_mitigated.yaml

run configs/tensortrust_extraction_25_gpt_oss_20b.yaml
run configs/tensortrust_extraction_25_gpt_oss_20b_mitigated.yaml

run configs/tensortrust_extraction_25_groq_compound.yaml
run configs/tensortrust_extraction_25_groq_compound_mitigated.yaml

# ── Group 4: HackAPrompt override (3 new models) ─────────────────────────────
run configs/hackaprompt_override_25_gpt_oss_120b.yaml
run configs/hackaprompt_override_25_gpt_oss_120b_mitigated.yaml

run configs/hackaprompt_override_25_gpt_oss_20b.yaml
run configs/hackaprompt_override_25_gpt_oss_20b_mitigated.yaml

run configs/hackaprompt_override_25_groq_compound.yaml
run configs/hackaprompt_override_25_groq_compound_mitigated.yaml

echo ""
echo "════════════════════════════════════════"
echo "  28 new experiments complete."
echo "════════════════════════════════════════"

# ── Benign control (all 7 models) ─────────────────────────────────────────────
run configs/benign_control_25.yaml
run configs/benign_control_25_mitigated.yaml

run configs/benign_control_25_llama70b.yaml
run configs/benign_control_25_llama70b_mitigated.yaml

run configs/benign_control_25_qwen32b.yaml
run configs/benign_control_25_qwen32b_mitigated.yaml

run configs/benign_control_25_llama4scout.yaml
run configs/benign_control_25_llama4scout_mitigated.yaml

run configs/benign_control_25_gpt_oss_120b.yaml
run configs/benign_control_25_gpt_oss_120b_mitigated.yaml

run configs/benign_control_25_gpt_oss_20b.yaml
run configs/benign_control_25_gpt_oss_20b_mitigated.yaml

run configs/benign_control_25_groq_compound.yaml
run configs/benign_control_25_groq_compound_mitigated.yaml

echo ""
echo "════════════════════════════════════════"
echo "  All 42 remaining experiments complete."
echo "════════════════════════════════════════"