Instructions to use lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B")
model = AutoModelForCausalLM.from_pretrained("lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B

SGLang

How to use lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B with Docker Model Runner:
```
docker model run hf.co/lalalamdbf/AgentSafeEvo-Defender-Qwen3-8B
```

AgentSafeEvo-Qwen3-8B-Defender

This is the defender in AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety. It is based on Qwen3-8B and trained to complete legitimate tool-use tasks while resisting harmful instructions from users and tools.

Training

The defender starts from the original Qwen3-8B model and is jointly trained with an evolving attacker using Group Relative Policy Optimization (GRPO). Its reward combines safety and legitimate task completion. Learnability-guided task weighting scales both policies' normalized advantages using within-task defender-reward variability.

Input and output

Input: system instructions, a user query, available tool definitions, and interaction history. Output: native tool calls and assistant responses across a multi-turn interaction.

Evaluation

Reported mean ± standard deviation over three runs. Lower ASR and harm score are better; higher UA, NRP, and accuracy are better.

Benchmark	Metric	Original Qwen3-8B	AgentSafeEvo defender
Agent-SafetyBench	ASR (%)	49.11 ± 0.51	17.84 ± 0.24
AgentHarm	Harm score	50.37 ± 3.63	6.53 ± 0.38
AgentDojo	ASR (%)	22.20 ± 0.57	6.18 ± 0.18
AgentDojo	UA (%)	53.64 ± 0.90	66.52 ± 1.43
MSB	ASR (%)	60.43 ± 0.97	31.25 ± 1.24
MSB	NRP	28.12 ± 1.30	46.00 ± 0.54
BFCL-V3	Accuracy (%)	66.13 ± 0.22	65.98 ± 0.14
ALFWorld	Accuracy (%)	30.28 ± 4.59	43.72 ± 3.03
ToolHop	Accuracy (%)	21.84 ± 1.23	48.04 ± 1.40

UA denotes Utility under Attack; NRP denotes Net Resilient Performance.

Intended use and limitations

Intended for agent safety research and tool-use evaluation. The model can still follow harmful instructions or fail legitimate tasks. Results depend on the prompts, tools, and evaluation setup; application runtimes remain responsible for tool permissions and execution.