AivexaNewsSearch
AI news for builders and product teamsChecked every hour

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

Collected Oct 14, 2026

US startup Artemis Security has launched Orion-1, a model built to investigate and respond to attacks on corporate networks on its own. Artemis says Orion-1 outperforms general-purpose models including Claude Opus 5.5 and GPT-6 Sol on its own benchmark. The model is live in a private preview for select customers on the Artemis platform.

Orion-1 ingests log data from several sources: authentication services, cloud setups, corporate endpoints, SaaS applications and network gear. It then links individual signals that would look harmless in isolation, such as a corporate account login, newly granted cloud permissions and a freshly created email inbox rule, and ties them back to a single attacker. Once it has assembled a picture of what happened, the model assigns a confidence score, locks in a verdict and recommends or kicks off a response.

Artemis argues that general-purpose models break down at this job. Fed billions of events, they lose track of the thread, according to the company, or they flag odd but harmless activity as an attack. That failure mode is expensive in a security operations center, where analysts already wade through noise and where a false positive can pull responders away from a real intrusion. This is the gap Artemis is aiming at: not raw hacking capability, but reliable judgement under a flood of telemetry.

Training leaned on Artemis' own security operations. The company fine-tuned Orion-1 on tasks drawn from that work, including spotting suspicious patterns in log data, hunting for known attack techniques, writing new detection rules and containing live attacks. Artemis says no customer data was used during training. It calls Orion-1 a Frontier Foundation Model, but its blog post describes only the fine-tuning, leaving it unclear whether the company built on top of an existing model from another provider. That ambiguity matters for buyers trying to judge what is genuinely novel here.

To measure the model, Artemis built a benchmark called Decision-Grade Readiness, or DGR. Most existing security tests quiz models on attack techniques or have them solve hacking puzzles. DGR instead 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, which pushes them toward showing their reasoning rather than guessing.

Artemis reports results across five tasks and tested Grok 4.7 and Kimi K3 alongside Opus 5.5 and GPT-6 Sol. The largest gap appears in attack reconstruction, where Orion-1 scored 80.8 against 58.3 for the next-best model. Artemis says GPT-6 Sol was the best comparison model. Only Orion-1 scores at or above the company's self-defined 80-point threshold across all five tasks. Two caveats are worth holding onto: the benchmark is Artemis' own, and two of the most capable models on the market, Anthropic's Mythos 5 and OpenAI's GPT-6 Astra, are absent from the comparison. With Opus 5.5, the results may largely reflect answers from the older Opus 4.8, since that model routes most cybersecurity tasks there.

On the Artemis platform, customers set limits per action type on how much the model can decide on its own. High-impact actions need human approval by default. That control scheme reflects a broader pattern in agentic security tooling: autonomy is granted action by action rather than all at once, so a mistaken verdict cannot immediately translate into a disruptive containment step like isolating a production host or revoking credentials.

Artemis only came out of stealth in April, disclosing $70 million from seed and Series A rounds roughly six months after the company was founded. It is a young company selling into a buyers' market where trust is the scarce commodity.

The backdrop is a tightening squeeze on cyber capabilities in frontier models. Providers keep locking down those skills because the same abilities help attackers. OpenAI considers its Astra model the first that could reach the highest cybersecurity risk tier, in which a model could develop and run cyberattacks on its own. Claude Opus 5.5 routes most cybersecurity tasks to the older Opus 4.8. In July, when Hugging Face tried to analyze a fully autonomous AI attack on its own infrastructure using commercial frontier models, their guardrails blocked analysis of the attack data entirely, and the team switched 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 hands out the less restricted GPT-5.5-Cyber only to verified defenders, and Anthropic takes a similar route, running Claude Mythos 5 through the Claude Security scanner and partner security products. The logic is that defenders get capability while attackers are kept out. In practice, verification gates create friction and exclude smaller teams, researchers and anyone responding to an incident outside a formal partnership. A likely trade-off of the restrictive approach is slower defense against fast-moving attacks, which is the gap Artemis and others are positioning against.

Why it matters: security teams that have watched general-purpose models refuse or fumble incident analysis now have a purpose-built option, though Orion-1 is limited to a private preview and its strongest evidence comes from a vendor-run benchmark. Developers building detection and response tooling should note the action-level permission model, since high-impact steps still require human sign-off and that shapes how such a system fits into an existing SOC workflow. A key question to watch, and this is inference, is whether independent testing reproduces the attack-reconstruction gap once models like Mythos 5 and GPT-6 Astra are included.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Artemis Security's new Orion-1 model is purpose-built for cyber defense and beats Claude Opus 5.5 and GPT-6 Sol on the company's own benchmark. The model fills a growing gap as general-purpose frontier models lock down their cyber capabilities, blocking defenders in the process. The article AI guardrails designed to stop hackers are now blocking defenders, and Orion-1 aims to fill the gap appeared first on The Decoder .