Picture a researcher working on how one family of viral vectors is recognized by human immune pathways. Under Anthropic’s usual safeguards, the AI tools she wants could refuse to help. On September 17, 2026, the company announced a beta program meant to fix that for verified scientists, and it has opened applications to the wider life science community.
It is called the Life Sciences Verification Program, or LSVP. Anthropic said it had already brought on dozens of organizations through early access. Access to its Mythos, Opus and Sonnet models comes with a refined set of safeguards that the company describes as more permissive for biology. The aim is to open up work blocked in its generally available Fable models: drug discovery, research biology, clinical development and manufacturing. Teams from academic labs to startups and pharmaceutical companies are the target. The beta starts with teams and institutions, and Anthropic plans to extend it to individual Pro and Max plans over time.
Two kinds of grant
Nobody walks straight in. Each applicant goes through a check of research credentials, security standards and ethical research oversight. Verified teams can then ask for one of two grants, usable in Claude Science, Claude.ai, Claude Code and the API.
A Standard Use grant covers most life science work and can be shared across an entire team for daily tasks. It renews once a year and applies to Mythos 5.1, Opus 5 and Sonnet 5, with future models added as they launch. Anthropic lists basic science, R&D, supply chain and manufacturing, clinical development, quality assurance, regulatory affairs, and investing and diligence among the uses.
High-risk Use is an add-on for work that Standard Use still blocks. It removes all safeguards that block life sciences requests, covers a single project instead of a team, and has to be renewed every six months. Anthropic’s example: a researcher with dual-use work holds one Standard Use grant for daily activity plus project-specific High-risk grants. At launch, those exist for Claude Opus 5 and Claude Sonnet 5. For Mythos they stay limited to a small set of entities with extra vetting, while Anthropic works with the US government to widen them. Other safeguards, such as the cyber classifiers, stay on.
Why it is built this way
Anthropic pointed to its September 2026 threat intelligence report, which describes increasingly sophisticated misuse attempts on its platform, some it judged could support biological weapons development. The company’s problem is that in biology a single request often looks the same whether the person is researching a viral pathogen to develop vaccines or trying to make a virus more transmissible. The worst scenarios, it said, are those where legitimate access is diverted or overtaken: through malware or account takeover, through rogue or coerced insiders, or through agents working in swarms or on long tasks that take dangerous actions nobody intended.
So the design leans on monitoring rather than refusal. Built with enterprise chief information security officers, it ties each organization’s access to the use cases it named in its application. Those descriptions should read like a job listing, with no sensitive information or intellectual property. Anthropic watches LSVP traffic for anything outside that scope and can flag incidents to the organization’s admins, who act within agreed timeframes.
Misuse, the company said, can be spread across many requests to look disconnected, which is why enforcement moves from blocking each request in real time to offline review of patterns. Legitimate work gets fewer interruptions. The price is retention: LSVP traffic is kept for 30 days. The data is strictly compartmentalized and cannot be used for model training or accessed by Anthropic’s life sciences research teams.
Limits at launch
LSVP runs in Anthropic’s first-party console for API use and in Claude for Enterprise and Team plans, not on third-party platforms. It is unavailable to organizations with a business associate agreement enabled, so customers holding protected health information need separate non-BAA organizations. In the API and Claude Science, users can switch between grants. In Claude.ai and Claude Code, only a preselected default grant applies at first.
Xaira said it will apply the models across its drug discovery engine, and Edison said the program lets it use them to discover and develop new medicines. Manifold Bio said it will use them safely within its engine. Anthropic expects hundreds of organizations to enroll in the first week.
