Part of Leviathan Platform · standalone license available

Agent Guardrails

A prompt-injection detector and a Python egress firewall, for AI agents and automation you build and run yourself. Screen what your tools read and write. Stop them phoning home to the wrong place. Machine identities already outnumber human ones 80:1 in the average enterprise (KPMG, 2026) — this is built for the agents nobody's watching.

$2,000/yr, per org — flat, unlimited seats and environments

What it actually catches

Three real ways an agent gets steered off-policy.

Injection

A scraped page tells your agent what to do

Indirect prompt injection: content your agent reads — a web page, a document, a memory record — carries an instruction it was never meant to follow.

Egress

Your own script phones home to the wrong place

Iron-Thread, a Python egress firewall built on sys.addaudithook, blocks unauthorized outbound connections and subprocess spawns before they leave your infrastructure.

Drift

What you already approved quietly changes

MCP tool and skill-file pinning: a hash-based check that flags a tool or skill's content changing after you first trusted it.

Three modules

Composable, zero dependencies.

Pricing

One flat license. No per-seat fees.

Card checkout: built and priced, temporarily paused during payment-processor identity verification. Reach out and we'll send payment instructions directly.