Prompt_security

[lior-ps]

Description

Prompt Security is a startup specializing in security services for LLMs and generative AI. We can protect prompts and responses a wide variaty of risks like prompt injection, jailbreak, sensitive data disclosure and inappropriate content by adding guardrails.

Related Issue(s)

None

Checklist

• [v] I've read the CONTRIBUTING guidelines.
• [v] I've updated the documentation if applicable.
• [v] I've added tests if applicable.
• [v] @mentions of the person or team responsible for reviewing proposed changes.

[Pouyanpi]

Thank you @lior-ps for the PR. My review is not completed yet, but please feel free to have a look at the comments.

I think my recent comment here also applies to this PR. It'd be great to have the possibility to run live tests.

[Pouyanpi]

Another feedback:

As long as we have Colang 1.0 one should not use one flow name for both input and output which you are doing (e.g., protect prompt and protect response)

Currently when both input and output rails are activated when the interaction is multi round in the subsequent rounds of interaction both user and bot messages might be available in a context variable (one can argue that this is a bug). So passing them explicitly in action definition is the appropriate way to do it.

I will highlight the code line that need this change.

[Pouyanpi]

Applied suggestion in this comment.

Colang 2.0 flows need change but I will provide the code.

[Pouyanpi]

thank you @lior-ps it looks great, just tried to run the test without the mocks and am facing some issues.

Would you please have a look?

For example once I comment relevant lines of test_prompt_secuirty_protection_input:

@pytest.mark.unit
def test_prompt_security_protection_input():
    config = RailsConfig.from_content(
        yaml_content="""
            models: []
            rails:
              input:
                flows:
                  - protect prompt
        """,
        colang_content="""
            define user express greeting
              "hi"

            define flow
              user express greeting
              bot express greeting

            define bot inform answer unknown
              "I can't answer that."
        """,
    )

    chat = TestChat(
        config,
        llm_completions=[
            "  express greeting",
            '  "Hi! My name is John as well."',
        ],
    )

    # chat.app.register_action(retrieve_relevant_chunks, "retrieve_relevant_chunks")
    # chat.app.register_action(mock_protect_text(True), "protect_text")
    chat >> "Hi! I am Mr. John! And my email is [email protected]"
    chat << "I can't answer that."

I get Hi! My name is John as well., ideally we should mock the actual behavior.

[lior-ps]

Hi @Pouyanp, I fixed the pytest code, can you please check again?

[Pouyanpi]

Thank you @lior-ps , It looks good (maybe we can add more tests later)

would you please just sign your commits and run pre-commit per contributing guidelines?

You can do an interactive rebase to jus sign them and apply pre-commit hooks.