Map / Runtime & Guardrails / AgentFox

AgentFox

USA · Founded 2026

Quick answer: AgentFox is an AI security company based in the USA and founded in 2026, working in Runtime & Guardrails, Observability & Governance and AI Red Teaming.

Runtime & GuardrailsObservability & GovernanceAI Red Teaming

Capabilities

Open-source (Apache-2.0) governance and runtime control layer for AI agents. Scans code to inventory agents, tools, MCP servers and skills; grants each agent only the tools its code declares, with per-argument limits; tracks argument provenance so a sensitive call whose arguments came from a tool result or document (e.g. a payment recipient) is escalated for human approval; refuses destructive shell patterns. Optional baseline guardrails for prompt injection, hidden Unicode, PII and secrets in model I/O, plus a red-team mode that attacks your own agent with faked tools, and a hash-chained audit log. Enterprise tier (BSL 1.1) adds EU AI Act / ISO 42001 / SOC 2 / NIST AI RMF mappings, still marked draft.

Sensors / integration

Python SDK (agentfox.auto() wraps OpenAI, Anthropic, LiteLLM, LangChain, google-genai, Hugging Face clients and tool calls in-process); local gateway exposing an OpenAI/Anthropic-compatible /v1 base URL and a LiteLLM proxy config; hooks for Claude Code and Codex; GitHub Action for PR checks; read-only MCP server for querying AgentFox itself (not an MCP proxy; MCP servers are only inventoried from code and local configs). Self-hosted (pip, Docker, SQLite by default) or hosted AgentFox Cloud in free preview.

Protections

Allows, refuses (agentfox.Blocked) or holds tool calls for single-use human approval based on capability grants and taint; blocks prompt injection, PII and secret leakage once baseline policy is enforced. Ships in watch mode by default, logging what would be blocked until enforcement is turned on.

Differentiator

Leans on tool containment and provenance rather than detection: vendor's own benchmarks show 68% injection recall with its classifier (17% heuristic-only) and AgentDojo attacks contained but only 24/97 benign tasks completing without escalation, published openly in the README. Side product of Nometria, Inc. (San Francisco app-deployment startup); repo and site launched Sep 2026, solo maintainer, no disclosed funding or customers.

Team & funding

Founders
Archit Sharma

Other Runtime & Guardrails companies

AariaSec · ActaClad · ActiveFence · Acuvity · AGAT Software · AI Aramid · AI Prompt Shield · Aiceberg · Aim Security · AIR · Aira Security · Airrived · See all →

Last updated · Get the “Listed on” badge