Arcjet Launches Custom AI Security Models to Protect AI Agents

New family of custom-trained models delivers fast, context-aware threat detection for AI chatbots, coding agents, and autonomous security agents

SAN FRANCISCO, Nov. 5, 2026 — Arcjet, the security runtime platform for AI agents, today announced a new family of custom AI security models designed to protect AI agents from attacks. The first model, Arcjet Himalia 1, is now generally available and powers prompt injection detection across Arcjet's platform. Two additional models, Arcjet Thebe 1 for coding agents and Arcjet Carme 1 for autonomous security agents, are available in early access.

AI agents routinely execute code, access sensitive data, and interact with external systems, making prompt injection a major security risk. Attackers can exploit these capabilities by embedding malicious instructions in web pages, emails, code repositories, and other content that agents encounter during legitimate tasks. Once processed, these instructions can redirect an agent's behavior, exposing credentials, modifying code, or triggering unauthorized actions.

Detecting these attacks is difficult because the same instructions can be malicious in one context and entirely legitimate in another. A shell command that appears suspicious in a customer support chatbot may be perfectly appropriate for a coding agent. Security agents present an even greater challenge because their work requires them to analyze vulnerability reports, malicious code, and attack techniques that conventional detection systems may incorrectly flag as threats.

Arcjet's new models address these challenges by providing specialized detection tailored to the environments in which agents operate. Rather than relying on large general-purpose language models to evaluate every request, Arcjet uses compact, fine-tuned classifiers that identify malicious instructions in milliseconds, enabling security decisions to be made before agents execute potentially harmful actions.

"Prompt injection looks very different depending on what an AI agent is supposed to do," said David Mytton, founder and CEO of Arcjet. "A coding agent needs to execute commands that would look suspicious in a chatbot, while a security agent spends its time reading about vulnerabilities and attacks. Applying the same detection model everywhere creates false alarms and misses real threats. We've built specialized models that understand those differences and can make security decisions fast and cheap enough to run on every request."

"NetBird includes an AI assistant that helps configure network environments, so Arcjet's prompt injection detection is an essential security layer for us. Arcjet's new models give us greater confidence in the security of our production environments," said Misha Bragin, CEO and co-founder of NetBird.

Purpose-built models for different AI agent environments

Arcjet's new model family includes three specialized detectors:

Arcjet Himalia 1: A general-purpose prompt injection detection model designed for chatbots, AI applications, and model endpoints. Available today, Himalia 1 identifies attempts to override application instructions, extract internal information, or manipulate agent behavior.

Arcjet Thebe 1: A specialized model for coding agents that distinguishes legitimate development activity from malicious instructions embedded in tool outputs, web pages, repositories, and other external content. Thebe 1 is designed to reduce false positives when agents execute commands or interact with code.

Arcjet Carme 1: A specialized model for autonomous security agents that analyze vulnerabilities, security advisories, and potentially malicious code. Carme 1 is designed to identify instructions that attempt to redirect an agent into leaking secrets, weakening security controls, or introducing malicious dependencies.

The models extend Arcjet's existing agent runtime security capabilities, which allow organizations to observe agent activity and detect shadow AI, enforce security policies, and maintain detailed audit records across AI workflows. Arcjet's detection integrates into coding agents and custom agents in-code, providing visibility into activity that may not pass through a centralized AI gateway.

Arcjet also combines specialized behavior classification with session-level analysis. Suspicious content can be evaluated against the agent's assigned task, tool activity, and session history to identify attacks that unfold across multiple interactions.

Faster detection with custom-trained security models

Arcjet developed its new models after identifying limitations in existing open-source prompt injection detectors, including inconsistent training data, licensing restrictions, and high rates of false positives and false negatives.

The company fine-tuned the models using manually reviewed public datasets and proprietary training data. The initial versions of its prompt injection models are based on Qwen3-0.6B, an open-weight model with 600 million parameters, and use a classification architecture that produces a security verdict without generating text.

In evaluations against two public prompt injection datasets, Arcjet Himalia 1 achieved F1 scores of 0.964 and 0.971, respectively. The model recorded precision scores of 96.9% and 97.8%, with recall scores of 95.8% and 96.4% across different eval datasets.

Arcjet also tests its models against encoded and obfuscated attacks, including base64, hexadecimal encoding, reversed text, and manipulated chat formatting. Deterministic decoders identify known attack patterns and allow the models to evaluate decoded content alongside the original input.

The company operates the models on infrastructure it controls and has developed a rapid retraining pipeline to incorporate new attack patterns and continually improve detection.

Availability

Arcjet Himalia 1 is available immediately to all customers using Arcjet's prompt injection detection through its SDKs and coding agent integrations.

Arcjet Thebe 1 and Arcjet Carme 1 are available in early access for organizations developing coding agents and autonomous security agents.

Developers can reach out to [email protected] to learn more or start a free trial at arcjet.com.

SOURCE Arcjet