AI guardrails designed to stop hackers are now blocking defenders, and Orion-1 aims to fill the gap

General-purpose AI models are becoming a growing problem for security teams, since the same skills also help attackers. Artemis Security wants to close that gap with a purpose-built model.

US startup Artemis Security has launched Orion-1, a model designed to investigate and respond to attacks on corporate networks on its own, beating Claude Opus 5.5 and GPT-6 Sol on the company's own benchmark. Orion-1 processes log data from auth services, cloud setups, corporate endpoints, SaaS apps, and network gear, then connects individual signals like a corporate account login, new cloud permissions, and a freshly created email inbox rule to tie them back to a single attacker. Once it has a picture of what happened, the model assigns a confidence score, locks in a verdict, and recommends or kicks off a response.

General-purpose models lose track when they're fed billions of events, according to Artemis, or they flag harmless but odd activity as an attack. Artemis fine-tuned Orion-1 on tasks drawn from its own security operations, including spotting suspicious patterns in log data, hunting for known attack techniques, writing new detection rules, and containing live attacks. No customer data was used during training, the company says. Artemis calls Orion-1 a "Frontier Foundation Model" but only describes this fine-tuning in its blog post, and it's unclear whether the company built on top of an existing model from another provider.

Artemis built its own benchmark to measure trust in security decisions

Artemis developed a benchmark called "Decision-Grade Readiness" (DGR) for the comparison. Most existing security tests quiz models on attack techniques or have them solve hacking puzzles, but DGR gives every model the same initial alert, the same environmental data, and the same tools a security team would have, then asks it to decide whether an attack is underway. The thousands of tasks all have known outcomes, and some contain no attack at all or include benign data designed to look malicious. Models are scored on both their verdict and the evidence they present.

Artemis reports results across five tasks and tested Grok 4.7 and Kimi K3 alongside Opus 5.5 and GPT-6 Sol. The biggest gap shows up in attack reconstruction, where Orion-1 scored 80.8 compared to 58.3 for the next-best model.

Orion-1 is live in a private preview for select customers on the Artemis platform, where customers set limits per action type on how much the model can decide on its own. High-impact actions need human approval by default. Artemis only came out of stealth in April, disclosing $70 million from seed and Series A rounds about six months after the company was founded.

Guardrails meant for attackers are blocking defenders too

Broad frontier models are getting harder for security teams to work with because providers keep locking down their cyber capabilities. OpenAI considers its Astra model the first that could reach the highest cybersecurity risk tier, where a model could develop and run cyberattacks on its own. Claude Opus 5.5 routes most cybersecurity tasks to the older Opus 4.8, and when Hugging Face tried in July to analyze a fully autonomous AI attack on its own infrastructure using commercial frontier models, their guardrails blocked the analysis of the attack data entirely. The team ended up switching to the open-weight model GLM 5.2 running on its own servers.

For defense, providers are shipping specialized models with restricted access instead. OpenAI only hands out the less restricted GPT-5.5-Cyber to verified defenders, and Anthropic takes a similar approach with its strongest model, running Claude Mythos 5 through the Claude Security scanner and partner security products. Both Mythos 5 and OpenAI's strongest model, GPT-6 Astra, are missing from the Orion-1 comparison, and with Opus 5.5, the benchmark results may largely reflect answers from the older Opus 4.8.