Releasing models responsibly
Model release practices: staged access, monitoring and abuse reporting, and what a small company can do
How the field thinks about releasing capable models, from staged release to structured access, and the parts of it that apply to a company like ours that serves models rather than training them.
Release is a safety decision
How a model reaches people decides what can be done about misuse afterwards. Weights published openly cannot be recalled or monitored. A model behind an API can be watched, limited and switched off. Most of the field's release practice is about choosing a point on that line, and about learning before moving along it.
Staged release
The best-known early example is OpenAI's GPT-2 in 2019, released in steps of increasing size so that risks and uses could be studied at each step. Solaiman and colleagues' report on that release set out the reasoning: time between stages lets researchers, the public and the developer see how a model is actually used before a more capable one follows.
Structured access
Toby Shevlane's 2022 paper named the wider idea structured access: people use the model through an interface the developer controls, at arm's length, so the developer can set what it does and stop users getting around those limits by modifying it. He identified cloud APIs as the setting where this works best.
The counterweight is the case for open weights. The US NTIA's July 2024 report concluded that the evidence then did not justify restricting open model weights, and recommended that government monitor the risks, collect evidence, and keep the option to act. Both positions agree on one thing: whoever controls access is the one who can monitor and respond.
Thresholds and frameworks
The largest labs now publish frameworks that tie release decisions to capability thresholds, cyber among them. OpenAI's Preparedness Framework names cybersecurity as a tracked category and says that at its Critical threshold it would stop further development until safeguards are specified. The Frontier Model Forum's February 2026 report on cyber thresholds found industry frameworks converging on two concerns: low-skilled actors gaining sophisticated capability, and autonomous end-to-end attacks. Anthropic's Claude Opus 5 system card describes classifiers on cyber use, with some vulnerability work allowed and some blocked, and an exemption programme for verified defenders. These are the labs' own descriptions of their own policies.
What applies to us
Senova Systems does not train frontier models. Klyroe routes requests to models we serve or rent, so our release decisions are about which model goes live in the product and what agents can do with it, and we hold the structured-access position by default: everything runs through our servers, where the controls in the overview article sit.
- A new model replaces the current one only after it scores better on our coding bench, and after our safety cases: it must refuse what the current model refuses and answer the benign look-alikes.
- Agents go live in stages: a draft, a test chat that must pass on that exact draft, a frozen published version, then deployment to a link, widget or API. Publishing and deploying each need a slide to approve, and a spending cap is set when the owner has none.
- Features that let agents act are behind switches that can turn them off for everyone, and new agent machines for coding sessions are off by default.
- Abuse reports: published apps carry a report form whose reasons include phishing and malware, and reports alert staff. Our acceptable use policy names the contact for abuse and security reports.
Not built yet: a staff screen for reviewing abuse reports (staff work from an internal endpoint and alert emails for now). Klyroe has no security.txt file yet; security reports go to [email protected] with the subject Security.
Monitoring without reading everyone's messages
Monitoring is where release practice meets privacy. The labs' threat reports show what monitoring at scale can find, and they come from reviewing activity on their platforms. A small company has to choose what it watches. Our abuse monitoring watches actions, not conversations: the activity log records tool calls with secrets removed, the audit log keeps no message content, and send limits count recipients as keyed hashes. That catches the shape of misuse (many sends, many new recipients, blocked connections) without a person reading private chats. It will miss misuse that never turns into an action, which is why the model's refusals and our evaluations stay part of the picture.
Sources
- Solaiman et al., Release Strategies and the Social Impacts of Language Models, arXiv, 2019
- Toby Shevlane, Structured access: an emerging paradigm for safe AI deployment, arXiv, 2022
- NTIA, Dual-Use Foundation Models with Widely Available Model Weights, 30 July 2024
- OpenAI, Preparedness Framework version 2, 15 April 2025 (lab's own framework)
- Frontier Model Forum, Managing Advanced Cyber Risks in Frontier AI Frameworks, 13 February 2026
- Anthropic, System Card: Claude Opus 5, July 2026 (lab's own report)
- Google DeepMind, Strengthening our Frontier Safety Framework, 22 September 2025