Humanbound

What it is

Humanbound is a security platform for AI agents that combines automated adversarial testing, runtime protection, and continuous monitoring. The testing engine runs black-box attacks against AI agent endpoints without requiring SDK installation or agent modification, adapting its attacks across single-turn, multi-turn, and agentic modes. The Humanbound Firewall sits between users and agents, evaluating every input through four defence tiers: sanitization, pre-trained attack detection, an agent-specific classifier, and LLM-based contextual analysis.

Every vulnerability gets mapped to frameworks including EU AI Act, NIST AI RMF, OWASP LLM Top 10, and OWASP Agentic AI Top 10, with severity ratings and reproducible evidence trails. The platform provides a posture score from 0 to 100 that updates continuously as agents, models and configurations change, rather than point-in-time reports.

The core testing engine, CLI, and firewall are open source under Apache-2.0 license with no login required. Users can run a complete security test locally using their own API keys or go fully air-gapped with self-hosted models like Ollama. The platform edition adds continuous monitoring, findings lifecycle management, cross-session intelligence, and managed infrastructure for teams securing multiple agents.

Questions from the bench

What types of AI agents can Humanbound test?

Any agent that exposes an API endpoint can be tested. Humanbound runs attacks as a black-box adversary without requiring SDK installation or agent modification. If a user can talk to the agent, Humanbound can test it.

How is this different from a traditional penetration test?

Humanbound provides a continuous posture score that updates as agents, models, and configurations change, rather than a point-in-time report. The testing engine adapts to agent behavior and runs multi-turn and agentic attack chains that evolve over time instead of replaying fixed payloads.

Can I run it fully air-gapped?

Yes. The local engine supports Ollama and other self-hosted models, allowing complete security tests without any data leaving the environment. No Humanbound account is required for local use.

How does the Humanbound Firewall work?

The firewall sits between users and the agent, evaluating every input before it reaches the model. Four tiers work together: input sanitisation, pre-trained attack detection, an agent-specific classifier trained on test data, and deep contextual analysis by an LLM judge. It ships as a Python package under Apache-2.0.

How long does a first assessment take?

A baseline campaign against a single agent endpoint typically completes in hours, not weeks. For an enterprise-wide assessment covering multiple agents, plan for about two weeks.