Skip to content

README.md

Latest commit

 

History

History
81 lines (63 loc) · 2.38 KB

README.md

File metadata and controls

81 lines (63 loc) · 2.38 KB

Trust-gated RAG agent

A LangChain agent that evaluates every external URL through the Entropy0 Trust Control Plane before fetching content from it.

Without a trust gate, RAG pipelines fetch from any source the LLM decides to visit — including newly registered domains, phishing infrastructure, and brand impersonation sites. This example adds one tool call between retrieval and synthesis that makes the decision explicit and auditable.

How it works

User query
    ↓
Agent decides to fetch a URL
    ↓
entropy0_trust_check(url)           ← Entropy0 /v1/decide
    ↓
recommended_action = "proceed"      → fetch + add to context
recommended_action = "proceed_with_caution" → fetch with reduced trust
recommended_action = "sandbox"      → block, inform user
recommended_action = "deny"         → block, inform user

The agent receives a plain-English result it can reason about:

Trust evaluation for secure-login-verify-account.xyz:
recommended_action=deny (confidence=91%), uncertainty=low,
signals=[NEWLY_REGISTERED_DOMAIN, BRAND_MISMATCH, CERTIFICATE_ANOMALY].

Setup

pip install -r requirements.txt

export ENTROPY0_API_KEY=sk_ent0_xxxx   # get from entropy0.ai/signup
export OPENAI_API_KEY=sk-...

Run

python agent.py

Expected output

Query: Summarise the content at http://secure-login-verify-account.xyz
...
> Entering new AgentExecutor chain...
> Invoking: entropy0_trust_check with {'target': 'http://secure-login-verify-account.xyz'}
  Trust evaluation: recommended_action=deny (confidence=91%), uncertainty=low,
  signals=[NEWLY_REGISTERED_DOMAIN, BRAND_MISMATCH, CERTIFICATE_ANOMALY].
...
Final answer: I was unable to fetch this URL. Entropy0 flagged it as a
high-confidence threat: newly registered domain with brand impersonation
signals and certificate anomalies. Recommended action: deny.

Customise

Change the policy to make the agent stricter or more permissive:

Entropy0Tool(
    api_key=os.environ["ENTROPY0_API_KEY"],
    policy="strict",          # stricter: sandboxes faster under ambiguity
    interaction_sensitivity="high",  # treats all fetches as high-sensitivity
)

Links