LLMs and biorisk

A Frontier Red Team post explaining why Anthropic treats biological weapons as a priority among AI misuse risks, citing Claude's rising scores on the Virology Capabilities Test, a bioweapons acquisition planning trial in which Claude 4 models raised participants' scores over an internet-only group, and a small 2024 wet-lab pilot that found no uplift; it then lists the safeguards Anthropic applies.

Anthropic (Frontier Red Team) · 2025-09-05 · original

This post asks why a developer of general-purpose language models should put biological weapons near the top of its list of misuse risks, and answers a series of questions "that might be posed by such a skeptic" (introduction). It was published on 5 September 2025 on the Frontier Red Team's blog, red.anthropic.com, under the label "Biorisk", and now appears on anthropic.com as "LLMs and biorisk" under the Frontier Red Team label with the same date and text; both versions carry the headline "Why do we take LLMs seriously as a potential source of biorisk?", and neither names individual authors. It came about three and a half months after Anthropic activated its AI Safety Level 3 (ASL-3) deployment protections for Claude Opus 4, a decision the post describes as "precautionary". The post is an explanatory essay that reports a few evaluation results; the evaluations themselves are documented in the Claude Opus 4 system card, which is the version of record for their numbers.

What the post argues

The post makes its case in five steps, each under a question heading.

  • Why AI matters for weapons at all. Threat actors have adopted earlier technologies (encrypted messaging, cryptocurrency, social media), and the post expects the same of AI. What it sees as new is that a model can act "as a true assistant": a guide to reliable information and to "otherwise tacit and inaccessible knowledge about implementation", which a static website cannot be (section "What does AI have to do with dangerous weapons at all?").
  • Why biology in particular. The post names two factors: the consequences of a successful biological attack can spread far beyond its first target, and falling costs of nucleic acid synthesis, standard reagent kits and common lab equipment have weakened material barriers that once acted as a "passive" biodefense (section "Why focus on biorisk?"). It notes that every published frontier safety framework it reviewed mentions biorisk, based on METR's collection of frameworks as of July 2025 (footnote 1).
  • Knowledge. Models are not yet able to do science autonomously at an expert level, the post says, but on several knowledge evaluations they are "nearing—and sometimes exceeding—expert human performance" (section "Don't experts know more about biology than AI models?").
  • Applied planning. A controlled trial compared participants with and without a model on a written planning task (next section).
  • The lab. Whether written knowledge translates into lab work depends on how much tacit knowledge lab biology requires, which a small pilot and a larger planned study address.

The post frames the subject as "low-probability/high-impact" and says its aim "is not alarmism" (introduction). A footnote adds that the size of the risk, "and even our ability to meaningfully estimate the magnitude", is "a source of healthy debate within Anthropic" (footnote 2).

What the evaluations found

The post reports three results. Its charts print no values; the numbers below come from the Claude Opus 4 system card where it gives them, and are marked where read from a chart.

  • Virology Capabilities Test (VCT). VCT is a multiple-choice evaluation of troubleshooting in virology lab scenarios, built by the nonprofit SecureBio. The post says that "within a year" Claude went from underperforming world-class experts to "comfortably exceeding that baseline" (Figure 1), and that the human baseline comes from scientists answering questions in their own specialty. The system card gives the expert baseline as 22.1%, scored on the hardest "multiple select" variant with 10-shot prompting, and calls the evaluation "saturated" because Claude Sonnet 3.7 already exceeds it (Claude Opus 4 system card, Section 7.2.4.4). Read from the post's Figure 1, which prints no values, scores rise from about 0.14 for Claude 3 Sonnet to about 0.40 for Claude Sonnet 3.7 and Claude Opus 4, with Claude Sonnet 4 lower at about 0.34; Claude 3 Opus already sits slightly above the baseline line (about 0.24). The post adds that models exceed expert performance "across multiple benchmarks, especially in molecular biology", without naming them.
  • Bioweapons acquisition planning trial. Participants had up to two days to write a plan for acquiring a biological weapon. A control group used basic internet resources; treatment groups also had Claude with safeguards removed. Biodefense experts graded the plans with a rubric covering key steps of the acquisition pathway. The post reports that participants with Claude 4 models, "especially Claude Opus 4", scored much higher and hit fewer critical failures than the control group (Figure 2, box plots of scores and of critical failures for control, Claude Sonnet 3.7, Claude Sonnet 4 and Claude Opus 4). The Claude Opus 4 system card describes a trial with the same design and the same box plots, and this note takes it to be the same trial (the post does not cite the card). The card gives the numbers: groups of 8 to 10 participants recruited through SepalAI and Mercor, grading by Deloitte, mean scores of 25% ± 13% for control, 42% ± 11% with Sonnet 4 and 63% ± 13% with Opus 4 (the card does not say whether ± is a standard deviation or an interval), an uplift of 1.70 times and 2.53 times; every participant hit critical failures (Section 7.2.4.1). The card's threat analysis set uplift of 2.8 times or less as acceptable and 5 times or more as significant additional risk, and concluded that Opus 4 was close enough to the line that ASL-3 could not be ruled out (Section 7.2.4.1).
  • Wet-lab pilot (2024). In a preliminary trial with basic biology protocols, some participants had a version of Claude and others only the internet. The post reports that the team "did not observe any evidence of uplift in this study", and that all participants, including the internet-only group, "did surprisingly well" in the lab. The study had eight participants ("a very small (n=8) experiment") and used "fairly basic" protocols; the post reads it as weak evidence that tacit knowledge "may not be as significant of a bottleneck as many have suggested" (section "Ok, but we've been talking about information...").

The post contrasts the planning trial with a 2023 RAND experiment of similar design, which found no statistically significant difference between plans made with language models and with the internet alone, and says frontier models "are now showing clearer warning signs" (footnote 3).

Safeguards and next steps described

The post lists measures rather than results (section "What can be done to address biorisk?"):

  • Information sharing with governments, including pre-deployment testing with the US Center for AI Standards and Innovation (CAISI) and the UK AI Security Institute, and support for transparency norms such as publishing a safety framework and dangerous-capability results.
  • Deployment safeguards based on Constitutional Classifiers, which monitor inputs and outputs and block "a narrow class of potentially harmful information". At the time these applied only to the Claude Opus 4 family. Platform-wide monitoring by the Safeguards team applies to all models.
  • A larger wet-lab study, co-sponsored through the Frontier Model Forum with the nonprofit Sentinel Bio, to test whether AI can bring non-experts to expert-level lab performance. The post reports no results from it.
  • National action: the post welcomes the biosecurity provisions of the July 2025 US AI Action Plan, including nucleic acid synthesis screening.

Instructions and texts involved

The post contains no principles or prompts. The one written text it relies on is the content policy behind the Constitutional Classifiers, which the post mentions only by name; the Constitutional Classifiers note describes how such a constitution of restricted and permitted content is used to generate training data.

Limitations

  • Proxies. The post itself calls text-based uplift trials "imperfect proxies for real-world scenarios" that leave out tacit knowledge, materials access and persistence (below Figure 2). The system card says "substantially weaker proxies" (Section 7.2.4.1).
  • Small samples and no published detail in the post. The planning trial used groups of 8 to 10 (system card), the wet-lab pilot eight participants; the post itself gives neither the trial's group sizes nor its rubric, and the rubric and the scored plans are not public.
  • Grading. The rubric grader was a contractor (Deloitte), not blinded by any procedure the post or card describes.
  • Benchmark saturation. VCT measures multiple-choice knowledge, and the system card already treats it as saturated, so further gains on it say little about risk.
  • Chart values. The VCT and uplift charts print no numbers, and the post's VCT chart starts with Claude 3 Sonnet; the claim of going from below to above experts "within a year" depends on which models are compared.
  • Self-report. All results come from evaluations Anthropic commissioned on its own models; no non-Anthropic model is tested.

Research questions

  1. Does the planning-trial uplift hold with larger groups and blinded grading? The card's control and Opus 4 ranges (25% ± 13% and 63% ± 13%) do not overlap, but rest on 8 to 10 people per group. A replication with larger samples and graders blind to condition would show how stable the 2.53 times figure is.
  2. How much does tacit knowledge limit novices in the lab? The 2024 pilot found the control group did well on basic protocols. The results of the Frontier Model Forum and Sentinel Bio study, with harder protocols and a larger sample, would answer this.
  3. What does VCT score predict once models pass the expert baseline? Correlating VCT scores with performance in planning trials or wet-lab tasks across model generations would show whether the benchmark carries information after saturation.
  4. How do critical failures change with model help? The post reports fewer critical failures with Opus 4 but every plan still had some. A breakdown by pathway step would show which barriers models lower and which remain.
  5. How often do the deployed classifiers block legitimate biology use? Published false-positive rates on real scientific traffic would show the cost of the safeguards the post describes.

How it connects

References