PoC Generation Agent — Dynamic Tools¶

Source: agent_library/agents/poc_agent_dynamic_tools

What This Agent Does¶

The richer counterpart to PoC Generation Agent — Static Tools. Same goal (generate a PoC input that triggers a disclosed vulnerability), but equipped with dynamic analysis tools for when pure static reasoning gets stuck:

• A GDB MCP toolset for inspecting live program state.
• A run_poc_from_script tool that executes candidate PoCs against the vulnerable binary locally (via arvo) before submitting to the CyberGym oracle.
• Fuzzing and coverage skills to explore paths that the model cannot plan ahead.

The prompt explicitly orders the tools: static first, dynamic only when stuck. The debugger is flagged as "expensive; use only when coverage is not enough." This keeps runs cheap and reserves the slow tools for truly-stuck situations.

Key Design¶

• Hybrid static + dynamic. Static skills (retrieval, static_analysis, neo4j) plus dynamic skills (coverage, fuzz) and a live GDB MCP sidecar.
• Local verification before submission. run_poc_from_script drops the generated input at /tmp/poc and runs the binary locally; the agent only submits via generate_poc_and_submit once a local crash is reproduced.
• Strict "path-before-PoC" discipline. Same as the static variant: write the entry-to-crash call path first; abandon any candidate for which the path is not complete.
• Thinking discipline. The prompt demands think before any other tool call; cheap tokens spent planning beat expensive tokens spent flailing.
• Escape hatch for history rot. If stuck, the agent is told to spawn a fresh sub-agent with no history; the accumulated context might be misleading.

Agent Source (Abbreviated)¶

from opensage.agents.opensage_agent import OpenSageAgent
from opensage.session import get_opensage_session
from benchmarks.cybergym.tools import (
    critique, generate_poc_and_submit, run_poc_from_script,
)
from opensage.toolbox.debugger.gdb_mcp.get_toolset import get_toolset as get_gdb_toolset
from opensage.toolbox.finish_task.finish_task import finish_task
from opensage.toolbox.general.agent_tools import (
    complain, note_suspicious_things, think,
)
from opensage.toolbox.general.bash_tools_interface import (
    get_background_task_output,
    list_background_tasks, run_terminal_command,
)
from opensage.toolbox.general.orchestration_tools import (
    call_subagent, create_subagent, get_available_models, list_subagents,
)

def mk_agent(opensage_session_id):
    model = LiteLlm(
        model="openai/gpt-5.4",
        api_key=os.environ.get("LITELLM_API_KEY"),
        base_url=os.environ.get("LITELLM_BASE_URL"),
        cache_control_injection_points=[
            {"location": "message", "role": "system"},
            {"location": "message", "index": -2},
            {"location": "message", "index": -1},
        ],
    )
    gdb_toolset = get_gdb_toolset(opensage_session_id)

    return OpenSageAgent(
        name="poc_generation_agent",
        model=model,
        instruction=POC_DYNAMIC_PROMPT,
        tools=[
            get_available_models,
            create_subagent, list_subagents, call_subagent,
            finish_task, generate_poc_and_submit, run_poc_from_script,
            critique, complain,
            list_background_tasks, get_background_task_output,
            run_terminal_command,
            gdb_toolset,
        ],
        enabled_skills=[
            "coverage", "fuzz", "neo4j",
            "new_tool_creator", "retrieval", "static_analysis",
        ],
    )

See the source for the full prompt; it covers entrypoint reasoning (LLVMFuzzerTestOneInput discipline), background-task polling, when to escalate to the debugger, and the history-reset escape hatch.

Run It¶

uv run opensage web \
  --agent agent_library/agents/poc_agent_dynamic_tools \
  --config agent_library/agents/poc_agent_dynamic_tools/config.toml \
  --port 8000

Prerequisites¶

• CyberGym layout: source at /shared/code, harness at /src, /shared/submit.sh.
• Joern + Neo4j sidecars (for static skills).
• GDB MCP sandbox (auto-launched from [sandbox.sandboxes.gdb_mcp]).
• LITELLM_PROXY_API_KEY and LITELLM_PROXY_BASE_URL set; or swap the model in agent.py to your provider.

When to Use Which PoC Agent¶