Security teams face a tricky problem with general-purpose AI: capabilities that can help investigate an attack can also help carry one out. Artemis Security says its new model, Orion-1, is designed for the defensive side of that challenge, by tracing suspicious activity across corporate systems and helping decide what to do next.
Connecting signals across company systems
Orion-1 processes logs from authentication services, cloud environments, corporate endpoints, software-as-a-service apps and network equipment. The aim is to make sense of events that might look unrelated when viewed in separate systems.
For example, the model could connect a corporate account login with newly granted cloud permissions and a recently created email inbox rule. Taken together, those clues may point to one attacker. After building a picture of events, Orion-1 assigns a confidence score, reaches a verdict and recommends a response or starts one.
Artemis says general-purpose models can lose track when given billions of events, or mistake unusual but harmless activity for an attack. It fine-tuned Orion-1 using tasks from its own security operations, including finding suspicious patterns in logs, searching for known attack techniques, creating detection rules and containing live attacks. The company says it did not use customer data in training.
A benchmark built around security decisions
Artemis created a test called Decision-Grade Readiness, or DGR, to assess how models handle security judgments. Rather than focusing only on knowledge of attack methods or hacking puzzles, the test gives each model the same alert, environment data and tools available to a security team. The model must decide whether an attack is underway.
The tasks have known outcomes, and some describe no attack at all. Others include harmless information intended to look suspicious. That design tests whether a model can distinguish a real threat from a false alarm, as well as explain the evidence behind its decision.
Artemis reports results across five tasks, comparing Orion-1 with Grok 4.7, Kimi K3, Claude Opus 5.5 and GPT-6 Sol. The largest reported gap was in attack reconstruction: Orion-1 scored 80.8, while the next-best model scored 58.3. These are results on Artemis's own benchmark, so they describe performance on that test rather than establishing how every model would perform in every security environment.
Why general-purpose models can be hard to use
AI providers have restricted some cybersecurity capabilities because powerful tools could enable attacks as well as defense. That creates a practical obstacle for defenders: a model's safeguards can also prevent it from analyzing threat data.
The source article describes a Hugging Face team that tried to analyze a fully autonomous AI attack on its own infrastructure using commercial frontier models. Their guardrails blocked analysis of the attack data, so the team switched to the open-weight model GLM 5.2 running on its own servers. The example illustrates the tension between limiting harmful use and allowing security teams to investigate real incidents.
Providers are also offering specialized models or access arrangements for defenders. The article says OpenAI limits GPT-5.5-Cyber to verified defenders, while Anthropic makes Claude Mythos 5 available through Claude Security scanner and partner security products. Those models, along with OpenAI's strongest model, GPT-6 Astra, were not included in Artemis's comparison. The article also notes that Claude Opus 5.5 routes most cybersecurity tasks to the older Opus 4.8, which may affect how to interpret the benchmark result.
Access and oversight remain part of the design
Orion-1 is available in a private preview for select customers on the Artemis platform. Customers can set limits on how much the model may decide independently for different action types, and high-impact actions require human approval by default. That lets teams define where automated assistance ends and a person must review the decision.
Artemis calls Orion-1 a "Frontier Foundation Model," but its blog post describes fine-tuning without making clear whether the company built on an existing model from another provider. That leaves an open question about the model's underlying foundation. For now, the benchmark offers one indication of how Orion-1 handles security decisions, while the private preview and approval controls show that deployment still depends on customer oversight.