Lakera Guard 2.0

Lakera's real-time guardrail API for LLM apps, updated with multi-agent protection and a new benchmark for detecting indirect prompt injection in tool outputs.

Lakera · Security and safety tooling · security

Published in Feed

Overview

Lakera released Guard 2.0 on October 14, 2026, adding an "Agent Shield" mode that screens tool-call outputs and sub-agent messages for injected instructions, not just end-user prompts, at a median added latency under 20 milliseconds per request. Alongside the release, Lakera published Gandalf Agent Bench, an open benchmark for measuring indirect prompt-injection resistance in multi-agent systems, built on red-team data collected from its public Gandalf game. Guard 2.0 ships drop-in SDKs for LangChain, LlamaIndex, and the Claude Agent SDK.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • Agent Shield mode screens tool-call outputs and sub-agent messages for injected instructions
  • Sub-20ms median added latency per screened request
  • Gandalf Agent Bench: an open benchmark for indirect prompt-injection resistance in multi-agent systems
  • Drop-in SDKs for LangChain, LlamaIndex, and the Claude Agent SDK
  • Continuously retrained on red-team data from Lakera's public Gandalf game

Use cases

  • Enterprises shipping customer-facing LLM agents that call external tools or APIs
  • Red-teaming and benchmarking agent robustness to prompt injection
  • Compliance teams needing audit logs of blocked attacks
  • Securing multi-agent pipelines where sub-agent messages are a new attack surface

Pricing

Free tier up to 10,000 requests/month; paid plans from $0.002 per screened request; Enterprise custom pricing.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
14 Oct 2026
Record updated
14 Oct 2026
Record last checked
14 Oct 2026

Explore the Feed

Other entries filed under security and safety tooling; inclusion does not imply equivalent features.

  • DefenseClaw Edge Connector

    Cisco's open-source DefenseClaw gains an Edge Connector that allows or blocks AI agent tool calls directly on IoT and robot hardware.

  • Fleming-1

    Sierra's model that listens to a live phone call and scores whether the caller is an AI agent rather than a human.

  • Anthropic Cyber Mission

    Anthropic's long-term program giving Claude models, engineers, and threat research to defenders of critical infrastructure and open-source software.

Browse the Feed