sologate-guard

Prompt firewall for AI agents.

Scores incoming prompts 0–100 before they reach your LLM — catches prompt injection, goal hijacking, credential fishing, and jailbreak attempts. Routes anything risky to the Sologate Decision Center for human review.

Install

npm install sologate-guard

Usage

Score a prompt (no gate)

import { scorePrompt } from 'sologate-guard'

const { score, tier, reason, flags } = scorePrompt(userInput)
// { score: 95, tier: 'HIGH', reason: 'Prompt injection — attempts to override agent instructions', flags: [...] }

Guard — wrap your LLM call

import { guard, GateRejectedError } from 'sologate-guard'

try {
  const response = await guard({
    prompt: userInput,
    sologateUrl: process.env.SOLOGATE_URL,
    apiKey: process.env.SOLOGATE_KEY,
    threshold: 70, // gate anything above this score (default: 70)
    call: () => openai.chat.completions.create({
      model: 'gpt-4o',
      messages: [{ role: 'user', content: userInput }],
    }),
  })

  // ✅ Prompt was safe (or human approved) — response is the LLM output
  console.log(response.choices[0].message.content)

} catch (err) {
  if (err instanceof GateRejectedError) {
    console.log('Prompt rejected by reviewer — aborting')
  }
}

Works with any LLM

// Anthropic
const response = await guard({
  prompt: userInput,
  sologateUrl: process.env.SOLOGATE_URL,
  apiKey: process.env.SOLOGATE_KEY,
  call: () => anthropic.messages.create({ ... }),
})

// LangChain
const response = await guard({
  prompt: userInput,
  sologateUrl: process.env.SOLOGATE_URL,
  apiKey: process.env.SOLOGATE_KEY,
  call: () => chain.invoke({ input: userInput }),
})

What it detects

| Pattern | Score | Default behavior | |---|---|---| | Prompt injection (ignore instructions) | 93–95 | Hard gate | | Persona override (you are now...) | 91 | Hard gate | | Goal hijacking (hidden secondary objective) | 90 | Hard gate | | Data exfiltration to external email | 97 | Hard gate | | Credential fishing (give me the API key) | 88 | Hard gate | | Jailbreak (DAN mode, no restrictions) | 92 | Hard gate | | Privilege escalation (run as root) | 75 | Gate (above threshold) | | Bulk data access (export all users) | 78 | Gate (above threshold) | | Social engineering (authority claim) | 72 | Gate (above threshold) | | Hypothetical harmful request | 65 | Gate (above threshold) | | Normal prompt | 10 | Auto-approved |

Environment variables

export SOLOGATE_URL=https://www.teamrocketlabs.com
export SOLOGATE_KEY=at_your_agent_key
export SOLOGATE_THRESHOLD=70   # optional — passed as threshold option

Pair with sologate-openclaw

sologate-guard watches the prompt (intent layer). sologate-openclaw watches the tool call (action layer).

Together they cover the full attack surface:

User prompt → [sologate-guard] → LLM → tool call → [sologate-openclaw] → execution

License

MIT