Claude Mythos Preview: procurement questions for an enterprise AI strategy

Do not build a business-critical plan around a model you have not been authorised to use. We recommend making three separate decisions: whether the task benefits from AI, whether a particular service is available on acceptable terms, and whether its implementation meets your operating requirements.

Announcement is not entitlement

Anthropic's 7 April 2026 Project Glasswing announcement described Claude Mythos Preview as an unreleased general-purpose model used with selected organisations for defensive work. It said there were no plans to make Mythos Preview generally available. That establishes the announcement's access posture, not a present right for every enterprise or implementation partner to use it.1

Its 22 May initial update reported progress and discussed future release considerations. It is a follow-up source, not a commitment that your organisation will receive model access on a particular date.20

This correction removes the earlier benchmark-label errors and any suggestion that security certifications are automatic access keys. Neither a coding score nor an industry certificate establishes a supplier's authorisation to provide a particular model to you.

A procurement evidence request

We suggest putting these questions to a proposed supplier:

Access: Which model, service, account and contractual route are actually available to this organisation? Ask for evidence of the entitlement, not a roadmap slide.
Architecture: Where do approved inputs, inference requests, outputs, logs and backups go? Which other providers or support teams can access them?
Permissions: What can the workflow read, prepare or change? What exact actions require approval, and what invalidates an approval?
Evaluation: Which permitted representative cases were used, which failed, and what was the human review burden? Separate vendor benchmarks from the proposed workflow's results.
Operation: Who monitors errors, authorises changes, handles incidents and verifies recovery? What happens when a model, account or dependency becomes unavailable?
Exit: Can you export the records, prompts, configuration and evidence needed to continue through another approved path? Specify what cannot be transferred.

The OAIC's commercial-AI guidance recommends due diligence on intended use, oversight, privacy and security risks, and who can access personal information. Take legal applicability questions to the appropriate adviser; a security architecture diagram is not clearance for every regulated activity.5

Keep the workflow independent of a forecast

Illustrative procurement exercise: a team wants a source-linked draft of a weekly operating brief. Compare the current manual method, available native tooling and a scoped workflow using a service the organisation can actually procure. Give each option the same approved inputs and acceptance conditions.

The business owner checks the brief before distribution. Missing information and conflicting numbers remain visible. If a model is unavailable, use the agreed manual process or a separately approved alternative, not an undisclosed fallback with different data handling. This is a recommended design, not an Amulet deployment claim.

What should change the decision?

We recommend a go/no-go record covering the task benefit, confirmed access, acceptable data flow, successful relevant evaluation, named owners, unresolved gaps and an exit path. A material change in model, permissions or data handling should return the affected decision to review.

You may not need an agent at all. Anthropic's engineering guidance recommends the simplest adequate solution and distinguishes predefined workflows from model-directed agents.10 That is a useful check against buying complexity for its own sake.

Amulet's public offer separates the AI Business BEEP diagnostic from a later implementation decision.17 For help defining the evidence you need before committing, discuss your workflow. No model entitlement, certification, residency or prebuilt connector is promised here.

Evidence checked on 15 September 2026; original publication date retained. Prepared with AI assistance. Amulet AI is responsible for this proposed correction. This is procurement analysis using dated vendor sources, not a prediction of model availability or a report of client results.

Sources

[1] https://www.anthropic.com/glasswing — Project Glasswing: Securing critical software for the AI era
[5] https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products — Guidance on privacy and the use of commercially available AI products
[10] https://www.anthropic.com/engineering/building-effective-agents — Building Effective AI Agents \ Anthropic
[17] https://amulet.ai/ai-business-beep — AI Business BEEP | Amulet AI
[20] https://www.anthropic.com/research/glasswing-initial-update — Project Glasswing: An initial update