A security startup's AI beats rivals on a test it wrote itself

Artemis Security, a US startup, has launched Orion-1, a model that investigates and responds to attacks on corporate networks, and says it outscores Claude Opus 5.5 and GPT-6 Sol at the job. The test behind that claim is Artemis’s own, and The Decoder’s report of the launch, published on 14 October, cites no independent evaluation.

Orion-1 ingests logs from login services, cloud setups, employee machines, software-as-a-service apps, and network equipment. It stitches scattered signals into one story: someone signs in to a staff account, cloud permissions change, an email forwarding rule appears. The model ties those to a single intruder, attaches a confidence score, commits to a verdict, and then proposes or launches a response. Customers cap how much it may do alone for each kind of action, and anything with a large impact waits for a human sign-off unless they change that.

What Orion-1 actually is remains thin. Artemis calls it a “Frontier Foundation Model,” yet its blog post describes only fine-tuning on tasks from the company’s own security operations, and The Decoder notes it is unclear whether Artemis built on another provider’s model. No size, no weights, and no price appear in the report. Only a handpicked set of customers on the Artemis platform can try it, in a private preview, and nobody outside the company is named as a user. The startup emerged from stealth in April and said then that it had raised $70 million across seed and Series A rounds, roughly six months after it was founded.

The benchmark deserves a fair description before it gets a skeptical one. Artemis calls it Decision-Grade Readiness. Across thousands of tasks with known answers, each model receives an identical opening alert and identical environment data, plus the toolkit an analyst would work with. Some scenarios involve no intrusion whatsoever, or harmless data engineered to look malicious, and scoring covers both the verdict and the evidence offered. But Artemis designed the test, tuned Orion-1 on its own operations tasks, and set the 80-point pass mark itself. Only Orion-1 clears it on all five tasks. The widest gap is attack reconstruction, 80.8 against 58.3 for GPT-6 Sol, the best rival.

The lineup has holes. Claude Mythos 5 and OpenAI’s strongest model, GPT-6 Astra, are absent. The Decoder adds that Opus 5.5 hands most of its security work to the older Opus 4.8, so the Claude result may largely reflect that older model.

Now the argument that makes the launch interesting. Artemis says general models lose the thread across billions of events and flag odd but harmless activity as attacks. That is a product claim, and Artemis is the only source for it. The claim that guardrails are blocking defenders rests on separate evidence in The Decoder’s reporting. OpenAI considers Astra the first model that could reach its highest cyber risk tier. And in July, Hugging Face set out to study an AI attack that ran with no human steering and hit its own systems. The commercial frontier models it tried refused to engage with the attack records at all. The team moved to GLM 5.2, an open-weight model, hosted on its own servers. That is one documented case from a defender with nothing to sell, not a survey of the field.

The labs are already building around the problem. According to The Decoder, OpenAI hands its looser GPT-5.5-Cyber variant only to verified defenders, and Anthropic runs Mythos 5 through its Claude Security scanner and partner products. Our own coverage adds two more. Anthropic merged its trusted-access schemes on 6 October into a Cyber Verification Program with three tiers, where qualifying applicants get its models with fewer automatic refusals. Google’s Fairwind Program gives vetted defenders a version of Gemini 4 Argon without the usual cyber guardrails, with no date for wider access.

So the gap is real enough that three of the biggest labs are vetting their way around it. What stays open is whether gated access is enough. Every gate is a lab’s decision about who counts as a defender, and Anthropic’s middle tier alone takes a few weeks to review. Hugging Face answered with open weights; Artemis answers with a vendor model. Neither has been tested head to head against the vetted options.

Until someone outside Artemis runs Orion-1 against the models the labs now hand to verified defenders, the 80.8 is a seller’s number, and the Hugging Face episode is the better evidence in this story.

Reported by The Decoder on 14 October 2026.