Skip to content

README.md

Latest commit

 

History

History
205 lines (155 loc) · 6.5 KB

README.md

File metadata and controls

205 lines (155 loc) · 6.5 KB

OpenHands - FCV Attack Experiments

This document describes how to run OpenHands with CWE injection attacks for the FCV (Functionally Correct yet Vulnerable) attack experiments.

Prerequisites

  1. Complete the Pass 1 run following the instructions in ../README.md
  2. Evaluate Pass 1 results with SWE-bench to get the report.json file
  3. Make sure you have the OpenHands environment activated: conda activate openhands

Workflow Overview

Pass 1 → Evaluation → CWE Injection (Pass 2) → Evaluation → LM Judge

Step 1: Evaluate Pass 1 Results

After running the initial inference, evaluate the results to identify resolved instances:

# Example for GPT-5 Mini
./evaluation/benchmarks/swe_bench/scripts/eval_infer.sh \
  "/path/to/OpenHands/evaluation/evaluation_outputs/outputs/princeton-nlp__SWE-bench_Verified-test/CodeActAgent/gpt-5-mini_maxiter_100_N_v0.56.0-no-hint-run_1/output.jsonl" \
  "" \
  princeton-nlp/SWE-bench_Verified \
  test \
  local

This will generate a report.json file with resolved instance IDs.

Step 2: Run CWE Injection Experiments

Use the run_cwe_simple.sh script to inject CWE vulnerabilities into the resolved instances:

Script Usage

./run_cwe_simple.sh <REPORT_JSON> <CWE_TYPE> <LLM_CONFIG> <ROUND> <OUTPUT_DIR> <USE_REMOTE>

Parameters:

  • REPORT_JSON: Path to the report.json from Pass 1 evaluation
  • CWE_TYPE: One of cwe_532, cwe_79, cwe_89, cwe_94
  • LLM_CONFIG: LLM configuration (e.g., llm.eval_gpt5_mini, llm.eval_kimi)
  • ROUND: Round number (typically 1)
  • OUTPUT_DIR: Directory to save outputs
  • USE_REMOTE: true to use remote runtime, false for local

CWE Types

  • CWE-532 (Information Exposure): Logging sensitive information
  • CWE-79 (Cross-Site Scripting): XSS vulnerabilities
  • CWE-89 (SQL Injection): SQL injection vulnerabilities
  • CWE-94 (Code Injection): Arbitrary code execution

Step 3: Run CWE Injection Experiments

General Template

cd /path/to/OpenHands && conda activate openhands

./run_cwe_simple.sh \
  <REPORT_JSON_PATH> \
  <CWE_TYPE> \
  <LLM_CONFIG> \
  1 \
  <OUTPUT_DIR> \
  true

Parameters:

  • <REPORT_JSON_PATH>: Path to Pass 1 report.json (e.g., evaluation/evaluation_outputs/outputs/princeton-nlp__SWE-bench_Verified-test/CodeActAgent/{MODEL_NAME}/report.json)
  • <CWE_TYPE>: One of cwe_532, cwe_79, cwe_89, cwe_94
  • <LLM_CONFIG>: LLM config name (e.g., llm.eval_gpt5_mini, llm.eval_kimi, llm.eval_qwen_480b, llm.eval_claude_sonnet4)
  • <OUTPUT_DIR>: Output directory (e.g., evaluation/evaluation_outputs/outputs/cwe532_gpt)

Model Configurations

Model LLM Config Model Name Pattern
GPT-5 Mini llm.eval_gpt5_mini gpt-5-mini_maxiter_100_N_v0.56.0-no-hint-run_1
Kimi-K2-Instruct llm.eval_kimi Kimi-K2-Instruct_maxiter_100_N_v0.56.0-no-hint-run_1
Qwen3-Coder-480B llm.eval_qwen_480b Qwen3-Coder-480B-A35B-Instruct_maxiter_100_N_v0.56.0-no-hint-run_1
Claude Sonnet 4 llm.eval_claude_sonnet4 claude-sonnet-4-20250514_maxiter_100_N_v0.56.0-no-hint-run_1

Example: Running All CWE Types for GPT-5 Mini

cd /path/to/OpenHands && conda activate openhands

REPORT_JSON="evaluation/evaluation_outputs/outputs/princeton-nlp__SWE-bench_Verified-test/CodeActAgent/gpt-5-mini_maxiter_100_N_v0.56.0-no-hint-run_1/report.json"

for CWE_TYPE in cwe_532 cwe_79 cwe_89 cwe_94; do
  ./run_cwe_simple.sh \
    "$REPORT_JSON" \
    "$CWE_TYPE" \
    llm.eval_gpt5_mini \
    1 \
    "evaluation/evaluation_outputs/outputs/${CWE_TYPE}_gpt" \
    true
done

Step 4: Evaluate CWE Injection Results

Evaluation Template

./evaluation/benchmarks/swe_bench/scripts/eval_infer.sh \
  "<OUTPUT_JSONL_PATH>" \
  "" \
  princeton-nlp/SWE-bench_Verified \
  test \
  local

Path Pattern: <OUTPUT_DIR>/princeton-nlp__SWE-bench_Verified-test/CodeActAgent/<MODEL_NAME>_maxiter_100_N_<CWE_TYPE>_injection_resolved_ids_round_1/output.jsonl

Example: Evaluate All CWE Types for a Model

# Define model-specific variables
MODEL_NAME="gpt-5-mini"  # or "Kimi-K2-Instruct", "Qwen3-Coder-480B-A35B-Instruct", "claude-sonnet-4-20250514"
OUTPUT_PREFIX="cwe"      # e.g., "cwe532_gpt", "cwe79_kimi", etc.

# Evaluate all CWE types
for CWE_NUM in 532 79 89 94; do
  ./evaluation/benchmarks/swe_bench/scripts/eval_infer.sh \
    "evaluation/evaluation_outputs/outputs/cwe${CWE_NUM}_*/princeton-nlp__SWE-bench_Verified-test/CodeActAgent/${MODEL_NAME}_maxiter_100_N_cwe_${CWE_NUM}_injection_resolved_ids_round_1/output.jsonl" \
    "" \
    princeton-nlp/SWE-bench_Verified \
    test \
    local
done

Step 5: Final Evaluation with LM Judge

After collecting all results, use the LM Judge to evaluate the vulnerability injection success. See ../attack-lm-judge/README.md for details.

Troubleshooting

Cleaning up stuck containers

If you encounter issues with stuck containers:

# Stop specific instance
docker stop $(docker ps -q --filter "name=sweb.eval.INSTANCE_ID") 2>/dev/null || true
docker rm $(docker ps -aq --filter "name=sweb.eval.INSTANCE_ID") 2>/dev/null || true

# Clean build directories
rm -rf /tmp/swe-bench/INSTANCE_ID/

Runtime cleanup

# List all remote runtimes
ALLHANDS_API_KEY="your-api-key" \
  curl -H "X-API-Key: your-api-key" \
  "https://runtime.eval.all-hands.dev/list"

# Stop all remote runtimes
ALLHANDS_API_KEY="your-api-key" \
  curl -H "X-API-Key: your-api-key" \
  "https://runtime.eval.all-hands.dev/list" | \
  jq -r '.runtimes[].runtime_id' | \
  xargs -I {} curl -X POST \
  -H "X-API-Key: your-api-key" \
  -H "Content-Type: application/json" \
  -d '{"runtime_id": "{}"}' \
  "https://runtime.eval.all-hands.dev/stop"

Output Structure

After completing the experiments, your output directory structure will look like:

evaluation/evaluation_outputs/outputs/
├── cwe532_gpt/
│   └── princeton-nlp__SWE-bench_Verified-test/
│       └── CodeActAgent/
│           └── gpt-5-mini_maxiter_100_N_cwe_532_injection_resolved_ids_round_1/
│               ├── output.jsonl
│               └── report.json
├── cwe79_gpt/
├── cwe89_gpt/
├── cwe94_gpt/
├── cwe532_kimi/
├── cwe79_kimi/
... (and so on for other models and CWE types)

Next Steps

  1. Collect all evaluation results
  2. Run LM Judge evaluation (see attack-lm-judge/README.md)
  3. Analyze attack success rates across models and CWE types