Anthropic hands vetted life science labs looser biology safeguards on its frontier models

Anthropic has opened applications for its Life Sciences Verification Program, a beta giving verified life science professionals access to Mythos, Opus and Sonnet models with safeguards the company describes as more permissive for biology-related work. Dozens of organisations are already in through early access, and Anthropic expects to enrol hundreds more in the first week. The tasks involved, drug discovery, research biology, clinical development and manufacturing, are currently blocked in the company's generally available Fable models. Until now, a legitimate lab and a bad actor met the same wall. The new wager is that verification, not blanket refusal, is the right answer for the legitimate ones.

Anthropic Watch analysis

What happened

Qualification starts with a verification process covering research credentials, security standards and ethical research oversight. Accepted teams then apply for grants tied to the use cases they declare, and those grants work across Claude Science, Claude.ai, Claude Code and the API. Launch partners including Xaira, Edison and Manifold Bio offered supportive quotes in the announcement.

What changes

• Standard Use grants
  Team-wide access to Mythos 5.1, Opus 5 and Sonnet 5 with classifiers more permissive for science tasks, renewed annually.
• High-risk Use grants
  An add-on that removes all life science safeguards for one named project, after extra vetting, renewed every six months.
• Offline monitoring
  Anthropic watches traffic against each organisation's declared scope and flags outliers to admins, rather than blocking requests in real time.
• 30-day data retention
  LSVP traffic data is kept for 30 days, compartmentalised, excluded from model training and off limits to Anthropic's life sciences teams.
• Where it works today
  The first-party API console plus Claude for Enterprise and Team plans. Individual plans and third-party platforms come later; BAA-covered health data stays out of scope in beta. Anthropic says it designed the safeguards around three threats: compromised accounts, rogue or coerced insiders, and agents, especially in swarms or on long-horizon tasks, going somewhere dangerous. Its recent threat report describes increasingly sophisticated misuse attempts, including ones that could support biological weapons development, and the company notes that insider threats and rogue use have featured in real biosafety incidents. High-risk grants exist for Opus 5 and Sonnet 5 today; for Mythos, Anthropic is working with the US government, and access stays with a small, extra-vetted set for now.

Why it matters

This is Anthropic, long among the strictest labs on biology requests, formalising a tiered loosening of its own guardrails. The bet is that verified organisations plus pattern monitoring beat blanket refusal, both for science and for safety, since real-time blocking cannot spot misuse deliberately spread across many sessions. If it works, expect the template to spread to other labs and other sensitive domains.

There is a privacy trade tucked inside. Thirty days of retained traffic on biology queries is a real change from ordinary expectations, compartmentalised or not. And declared use cases become the yardstick, so teams should scope grants carefully: that scope is what they will be measured against.

Our read

This reads as considered safety engineering rather than a guardrail fire sale. Tight scopes, tiered vetting, admin accountability, and the honesty to name insiders and agent swarms as the threat. The residue is time. Offline monitoring sees patterns after the requests, and in biology 'after' can be too late. Life science readers should apply, keep grant scopes narrow, and treat the High-risk tier with the respect its six-month renewals imply. Whether the monitoring has teeth is the part only incidents will test.

What to watch

• Mythos high-risk access
  Whether the US government work widens those grants beyond the initial vetted few.
• Enrolment pace
  Whether hundreds of organisations genuinely clear verification in week one.
• Individual plans
  When Pro and Max users get access, and what personal vetting looks like.
• The first flag
  Whether offline monitoring publicly catches a real misuse case, which would make or break the model.

Discussion spark: Anthropic is swapping real-time blocking for after-the-fact monitoring in biology. Is 30 days of retained traffic a fair price for unblocked science, or does offline detection only ever arrive after the harm?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.