Artemis Launches Orion-1, Defense AI Tested at Home DGR

Artemis Security highlights Orion-1, an AI model trained on its own security tasks and evaluated on an internal benchmark. Its claimed scores surpass those of generalist models, as access to the cyber capabilities of large models tightens. The tool is currently in private preview with operational safeguards on the client side.

Absences and Gray Areas in Comparisons

Two powerful models mentioned in the ecosystem, Claude Mythos 5 and GPT-6 Astra, are not included in the comparison published with Orion-1. With Opus 5.5, the results may largely reflect responses from the previous Opus 4.8. Furthermore, Artemis describes Orion-1 as a Frontier Foundation Model but only details this optimization in a blog post, and it is not established whether it is based on a model from another provider.

Operational Functions and Human Oversight of Orion-1

Orion-1 centralizes logs from authentication services, cloud configurations, endpoints, SaaS applications, and network equipment. It associates signals such as logging into a corporate account, granting new cloud permissions, or creating an email rule with the same attacker, then assigns a trust score, confirms a verdict, and proposes or initiates a response. The tool is accessible in private preview on the Artemis platform, where users set thresholds by action category for automatic decisions, with high-impact actions always requiring human validation.

How Artemis Claims to Measure: the DGR

Artemis is behind the Decision-Grade Readiness benchmark, which places each model against an identical initial alert, with the same contextual information and tools as a security team, and asks it to determine if an attack is underway. The DGR, consisting of thousands of tasks with known outcomes, some of which do not involve an attack or are intentionally misleading but harmless, evaluates both the verdict and the justifications provided. According to the results reported by Artemis on five tasks, only Orion-1 reaches or exceeds the threshold of 80 points, scoring 80.8 points on attack reconstruction, compared to 58.3 for the best comparison model.

Claimed Training and Limitations of Generalist Models

Artemis argues that generalist models lose track when faced with billions of events or misinterpret unusual but harmless activities as attacks. The company states it has refined Orion-1 on tasks drawn from its security operations, including detecting suspicious patterns, searching for known techniques, writing rules, and managing ongoing incidents, without using customer data. Artemis presents Orion-1 as a response to the ambivalence of large models' capabilities, which can also be useful to attackers.

Safeguards That Also Hinder Defense Analysis

The increasing restrictions imposed on the cyber capabilities of generalist models complicate their use by security teams. OpenAI describes Astra as the first model capable of reaching the highest risk level in cybersecurity, where an AI would autonomously develop and execute attacks. Claude Opus 5.5 redirects most tasks to the previous Opus 4.8. In July, an attempt by Hugging Face to analyze a fully autonomous attack with commercial models was blocked by their safeguards, leading the team to turn to GLM 5.2 on its own servers. On the defense side, providers favor specialized and restricted-access models, such as GPT-5.5-Cyber distributed to verified defenders at OpenAI, or Claude Mythos 5 operated via the Claude security scanner and partners at Anthropic.

Commercial Status and Funding of Artemis

Artemis Security launched Orion-1, designed to autonomously investigate and respond to cyberattacks on corporate networks. The company emerged from the shadows in April by announcing $70 million from seed and Series A rounds, about six months after its founding.