agentshield

Runtime security middleware for LLM agents. Drop it into your existing agent framework — no refactor required.

Defends against:

• Prompt injection (user input hijacking system instructions)
• Tool misuse & privilege escalation (agents calling tools outside intended scope)
• Memory poisoning (persistent agent memory corrupted across sessions)
• RAG context injection (poisoned document chunks injecting instructions)

Install

pip install apexguard
# or
npm install @apexguard/sdk

Python — 30-second quickstart

from apexguard import Shield, Policy

shield = Shield(
    policy=Policy(
        tool_allowlist=["calculator", "web_search"],
        memory_ttl=3600,
        injection_sensitivity="medium",
        on_violation="block",
    )
)

# Wrap your existing LangChain agent
secured_agent = shield.wrap(your_langchain_agent)
result = secured_agent.invoke({"input": user_query})

LangChain tool integration

from apexguard.adapters.langchain import shield_tools

secured_tools = shield_tools(your_tools, shield)
agent = create_react_agent(llm, secured_tools, prompt)

Custom violation handler

shield.on_violation(lambda e: send_to_slack(e))

TypeScript — 30-second quickstart

import { Shield } from '@apexguard/sdk';

const s = new Shield({
  policy: {
    toolAllowlist: ['calculator', 'webSearch'],
    memoryTTL: 3600,
    injectionSensitivity: 'medium',
    onViolation: 'block',
  },
});

const securedAgent = s.wrap(yourAgent);

Vercel AI SDK tools

import { shieldTools } from '@apexguard/sdk/adapters/vercel-ai';

const result = await generateText({
  model: openai('gpt-4o'),
  tools: shieldTools({ calculator, webSearch }, s),
  prompt: userInput,
});

Modules

Policy options

License

Apache 2.0