ai-risk-extractor

ai-risk-extractor is a lightweight Python package that extracts structured risk insights from free‑form user prompts about autonomous AI agents, task injection, AI agency, and related vulnerabilities. By leveraging a language model (default ChatLLM7), the package parses the input text and returns a standardized, machine‑readable summary that highlights threat levels, involved components, and possible exploitation methods.

Installation

pip install ai_risk_extractor

Quick Start

from ai_risk_extractor import ai_risk_extractor

# Example user prompt describing an AI risk scenario
prompt = """
An autonomous AI assistant receives a hidden instruction from a malicious user
that causes it to execute a privileged system command. The instruction is
embedded in a seemingly harmless chat message.
"""

# Extract structured risk information (uses default ChatLLM7)
risk_summary = ai_risk_extractor(user_input=prompt)

print(risk_summary)

Function Signature

def ai_risk_extractor(
    user_input: str,
    api_key: Optional[str] = None,
    llm: Optional[BaseChatModel] = None,
) -> List[str]:
    """
    Process `user_input` with a language model and return extracted risk data.

    Parameters
    ----------
    user_input: str
        The free‑form text describing AI scenarios or concerns.
    api_key: Optional[str]
        API key for the default `ChatLLM7`. If omitted, the function will
        read the `LLM7_API_KEY` environment variable. If that is also missing,
        a placeholder key `"None"` is used (the request will still be routed
        to the LLM7 endpoint).
    llm: Optional[BaseChatModel]
        Any LangChain `BaseChatModel` instance. If omitted, `ChatLLM7` from
        `langchain_llm7` is instantiated automatically.

    Returns
    -------
    List[str]
        A list of extracted data strings that match the internal regex pattern.
    """

Using a Custom LLM

You can provide any LangChain‑compatible chat model instead of the default ChatLLM7.

OpenAI (e.g., GPT‑4)

from langchain_openai import ChatOpenAI
from ai_risk_extractor import ai_risk_extractor

my_llm = ChatOpenAI(model="gpt-4")
result = ai_risk_extractor(user_input=prompt, llm=my_llm)

Anthropic (e.g., Claude)

from langchain_anthropic import ChatAnthropic
from ai_risk_extractor import ai_risk_extractor

my_llm = ChatAnthropic(model="claude-2.1")
result = ai_risk_extractor(user_input=prompt, llm=my_llm)

Google Generative AI

from langchain_google_genai import ChatGoogleGenerativeAI
from ai_risk_extractor import ai_risk_extractor

my_llm = ChatGoogleGenerativeAI(model="gemini-1.5-pro")
result = ai_risk_extractor(user_input=prompt, llm=my_llm)

API Key & Rate Limits

The default ChatLLM7 free tier provides generous rate limits suitable for most development and research workflows. If you need higher limits, obtain a personal API key by registering at:

https://token.llm7.io/

Provide the key either:

• via the LLM7_API_KEY environment variable, or
• directly as the api_key argument:

result = ai_risk_extractor(user_input=prompt, api_key="YOUR_LLM7_API_KEY")

Contributing

Contributions, suggestions, and bug reports are welcome! Feel free to open a pull request or discuss enhancements.

Issues

If you encounter any problems, please open an issue on GitHub:

https://github.com/chigwell/ai_risk_extractor/issues

License

This project is licensed under the MIT License.

Author

Eugene Evstafev
Email: hi@euegne.plus
GitHub: @chigwell