PoC Generation Agent — Static Tools¶

Source: agent_library/agents/poc_agent_static_tools

What This Agent Does¶

Given a vulnerability description and the source code of a vulnerable program, this agent generates a proof-of-concept (PoC) input that reproduces the crash. It is the OpenSage team's baseline PoC agent for the CyberGym benchmark using a static-only tool palette.

The premise: most PoCs can be derived by reasoning carefully about the code path from the fuzzer harness (LLVMFuzzerTestOneInput) to the vulnerable function, without ever running the program locally. The agent submits candidate PoCs to the CyberGym server and iterates based on the oracle's response.

For the dynamic counterpart (that adds GDB, fuzzing, coverage) see PoC Generation Agent — Dynamic Tools.

Key Design¶

• Static-first philosophy. No debugger, no fuzzer, no local execution. The prompt requires the agent to write an ordered, end-to-end path from entrypoint to crash before writing a PoC; if the path cannot be constructed, the candidate vulnerability is abandoned.
• Skill-backed retrieval. enabled_skills=["new_tool_creator", "retrieval", "static_analysis", "neo4j"] gives the agent access to bash-tool scripts (read-file, list-functions-in-file, search-symbol-definition, get-call-paths-to-function, joern-query, neo4j-query). The prompt explicitly prefers these over raw grep/sed.
• Dynamic sub-agents mandatory. The prompt promotes dynamic sub-agents to "preferred and default"; any subtask that can be broken down must go through create_subagent + call_subagent.
• Submission via CyberGym oracle. generate_poc_and_submit pushes a candidate PoC to the CyberGym server; /shared/submit.sh is the shell equivalent.

Agent Source (Abbreviated)¶

from opensage.agents.opensage_agent import OpenSageAgent
from benchmarks.cybergym.tools import (
    generate_poc_and_submit, run_poc_from_script,
)
from opensage.toolbox.finish_task.finish_task import finish_task
from opensage.toolbox.general.agent_tools import (
    critique, note_suspicious_things,
)
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-5.5",
        api_key=os.environ.get("LITELLM_API_KEY"),
        base_url=os.environ.get("LITELLM_BASE_URL"),
        reasoning_effort="medium",
    )
    return OpenSageAgent(
        name="poc_generation_agent",
        model=model,
        instruction=POC_STATIC_PROMPT,
        tools=[
            get_available_models,
            finish_task, generate_poc_and_submit,
            create_subagent, list_subagents, call_subagent,
            critique,
            list_background_tasks, get_background_task_output,
            run_terminal_command,
        ],
        enabled_skills=["new_tool_creator", "retrieval", "static_analysis", "neo4j"],
    )

See the full prompt for the end-to-end reasoning template and the "Dynamic Agent Usage (Very Important)" rules.

Run It¶

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

Prerequisites¶

• CyberGym-compatible target layout: source at /shared/code, harness at /src, /shared/submit.sh present.
• Joern + Neo4j sidecars, declared in the config.
• LITELLM_PROXY_API_KEY and LITELLM_PROXY_BASE_URL if you keep the default litellm_proxy/sage-gpt-5 model; otherwise swap LiteLlm(model=...) to your provider.