Anthropic loosens Claude's biology safeguards for vetted teams, removes them for some projects
Anthropic now offers verified life science teams a version of Claude with looser biology safeguards, and for approved high-risk projects it removes every safeguard that blocks life science requests. The Life Sciences Verification Program, announced on September 17, is in beta for organisations on Claude Enterprise, Team and the API, and not yet for anyone on a personal Pro or Max plan.
Claude running quality control on single-cell RNA data, from Anthropic's October 2025 Claude for Life Sciences announcement. Image: Anthropic
Anthropic says the programme covers work that is currently blocked in its generally available Fable models, such as drug discovery, research biology, clinical development and manufacturing. It has onboarded dozens of organisations through an early-access phase and now takes applications from the wider life science community.
Two grants decide how much Claude will answer
Every applicant goes through a verification process that reviews research credentials, security standards and ethical research oversight. Verified teams can then apply for one of two grants.
Standard Use grants cover whole teams and renew once a year. They unlock Mythos 5.1, Opus 5 and Sonnet 5 "with refined classifiers that are more permissive for science tasks than our generally available models", and Anthropic expects them to cover most life science work.
High-risk Use is the unusual one. Anthropic describes it as "an add-on grant for teams working in areas blocked under Standard Use" and says plainly: "It removes all safeguards that block life sciences requests."
The grant has limits. It covers only safeguards that block life science requests, applies to a single research project rather than a whole team, and it must be renewed every six months. Anthropic's example is a researcher characterising how one family of viral vectors is recognised by human immune pathways.
High-risk grants are available for Claude Opus 5 and Claude Sonnet 5. For Claude Mythos, Anthropic says it is working with the US government to widen access, and for now the grants stay limited to a small set of entities with extra vetting. Other safeguards, including Claude's cyber classifiers, stay in place under every grant.
Anthropic swaps live blocking for monitoring
Biology is the hard case for AI safeguards because, as Anthropic notes, the same request can come from someone developing a vaccine or someone trying to make a virus spread faster. Its answer is to vet the organisation rather than judge each prompt.
For programme traffic, Anthropic moves from rejecting risky requests in real time to reviewing usage patterns afterwards. Each organisation's access is tied to the use cases it stated in its application, and Anthropic says it can flag activity outside that scope to the organisation's admins, who then act within agreed timeframes.
The trade-off is data retention. Traffic under the programme is kept for 30 days for this monitoring, and Anthropic says that data is compartmentalised, cannot be used for model training, and cannot be accessed by its own life sciences research teams.
The threats it names are compromised accounts, insiders who go rogue or are coerced, and agents that take unintended dangerous actions on long tasks. Vetting an organisation does not stop any of them, which is the case Anthropic makes for monitoring usage against each stated use case.
Individual plans have to wait
For now, the programme runs in Anthropic's own console for API use and in Claude for Enterprise and Team. It is not available on third-party platforms, and as a beta it is not available to organisations with a HIPAA business associate agreement, which Anthropic says should use separate organisations for patient data.
Anthropic says it plans to expand access to individual Pro and Max plans over time, and expects to enrol hundreds of organisations in the first week. Life science companies Xaira Therapeutics, Edison Scientific and Manifold Bio are quoted in the announcement. No pricing is given.
Tiered access for dangerous capabilities is spreading. OpenAI takes a similar route by limiting its GPT-5.6 Cyber model to approved partners. Anthropic's version goes a step further for biology: for an approved project, none of the safeguards that block life science requests apply, though its cyber classifiers stay on.
