Anthropic opens verified biology access with fewer blocks and monitoring
Life science teams can apply for more permissive model access. The beta exchanges some request-by-request blocking for monitoring, with mandatory data retention.
The briefing
Anthropic opens a verified route to fewer biology restrictions, with monitoring and data-retention conditions attached. Google DeepMind’s new institute puts forward a model-testing proposal, while Microsoft opens its AI code of conduct to public feedback. Access is changing; the proposed industry rules are not yet rules.
Anthropic opens verified biology access with fewer blocks and monitoring
Anthropic’s Life Sciences Verification Program offers vetted teams fewer biology restrictions across its models. Access depends on approved uses, monitoring and renewal, with tighter conditions for higher-risk projects.
Anthropic is accepting applications from institutions and teams after onboarding dozens of organisations through early access. Applicants face checks on research credentials, security standards and ethical oversight.
Standard Use grants cover broad life science work and renew annually. They apply to Mythos 5.1, Opus 5 and Sonnet 5. Project-specific High-risk Use grants remove all safeguards that block life sciences requests and renew every six months; other protections, including cyber safeguards, remain.
Access comes with oversight
High-risk access is available for Opus 5 and Sonnet 5. Mythos access remains limited to a small, additionally vetted group while Anthropic works with the US government on broader availability.
Anthropic monitors activity against each grant’s declared uses and can flag unauthorised activity to administrators. LSVP traffic requires 30-day data retention. The company says this data cannot be used for training or accessed by its life sciences research teams.
The beta is available through Anthropic’s first-party API console and Claude Enterprise and Team plans, not individual plans or third-party platforms. Organisations enabled for Business Associate Agreements are also excluded from the beta.
Labs can seek access for biology tasks that ordinary safeguards block, but must also arrange oversight of approved uses and accept data retention. A grant changes the model’s access rules, not the need to validate its scientific answers.
Market signal
DeepMind institute opens with proposal for US frontier model tests
Google and Google DeepMind researchers have launched an institute publishing competing views on advanced AI. Its first essays include proposals for pre-release evaluations and limits on reasoning that humans cannot inspect.
The DeepMind Institute begins with four essays covering economic policy, readable model reasoning, human flourishing and frontier-model evaluation. Aditya Mehta reported the launch for TechCrunch.
Its directors are DeepMind co-founder Shane Legg, Google executive James Manyika and Google DeepMind chair Demis Hassabis. Legg also serves as managing editor. The institute aims to publish differing views, rather than a single agreed position.
Voluntary first, potentially compulsory later
Hassabis proposes a US-led standards body to assess the most advanced models. Developers would initially submit them voluntarily, up to 30 days before release. Once the system had proved effective, passing its tests could become a condition of deployment in the United States.
Assessments would initially be designed with AI companies. The body would later develop independent, undisclosed tests to stop developers tailoring models to known examinations. Hassabis also leaves room for a coordinated slowdown if the risks warrant it.
Keeping reasoning inspectable
In a separate essay, safety researchers Rohin Shah and Anca Dragan propose confronting the trade-offs of less transparent models. Options include limiting sequential computation without readable reasoning, or requiring evidence that opaque systems remain equally monitorable. These are proposals, not adopted requirements.
If adopted, the testing framework would add an external assessment before frontier-model deployment. Its move towards undisclosed tests addresses a specific weakness: a model can perform well on familiar evaluations without demonstrating equally reliable behaviour elsewhere.
What to watch
Microsoft opens AI conduct code to six weeks of feedback
Microsoft has published a 37-page Humanist AI Code of Conduct for public consultation. Its AI chief argues that training models to follow instructions must be paired with containment and independent scrutiny.
Microsoft AI chief Mustafa Suleyman set out how the company intends to use its new code in an interview with Nilay Patel of The Verge. The public consultation will remain open for six weeks, he said.
The code is intended to govern model development, not simply describe company values. Suleyman said Microsoft uses it to derive training data, create safety guardrails and evaluate models in use, rather than feeding the document into training unchanged.
Instructions are only part of control
Suleyman argues that better instruction-following does not remove the need to contain a model’s actions. His proposed approach combines alignment, meaning behaviour shaped towards human objectives, with limits on agency and the systems a model can reach.
One concrete position is that agents should communicate in human language rather than direct mathematical representations or opaque codes. That would give auditors something they can inspect, although Suleyman acknowledged that the volume of communication would still be difficult to oversee.
He also called for industry standards and independent third-party verification. He said lab leaders broadly shared the concerns, but had not agreed the precise measures. Microsoft’s consultation does not establish an industry-wide enforcement system.
The consultation gives outsiders a chance to challenge the principles Microsoft says will shape its models. For teams evaluating AI agents, it also separates two questions: whether a model follows instructions, and whether its permitted actions are safely bounded.
What to watch next
- Labs can seek access for biology tasks that ordinary safeguards block, but must also arrange oversight of approved uses and accept data retention. A grant changes the model’s access rules, not the need to validate its scientific answers.
- If adopted, the testing framework would add an external assessment before frontier-model deployment. Its move towards undisclosed tests addresses a specific weakness: a model can perform well on familiar evaluations without demonstrating equally reliable behaviour elsewhere.
- The consultation gives outsiders a chance to challenge the principles Microsoft says will shape its models. For teams evaluating AI agents, it also separates two questions: whether a model follows instructions, and whether its permitted actions are safely bounded.
The takeaway
Labs can seek access for biology tasks that ordinary safeguards block, but must also arrange oversight of approved uses and accept data retention. A grant changes the model’s access rules, not the need to validate its scientific answers.
The editor’s view
If adopted, the testing framework would add an external assessment before frontier-model deployment. Its move towards undisclosed tests addresses a specific weakness: a model can perform well on familiar evaluations without demonstrating equally reliable behaviour elsewhere.