{"id":"1594a8990a7e1d707e9ca08b0aaa1dc05e01ac6eb6c6b4a6849ee2738ef14e8e","title":"Anthropic argues for treating AI biorisk as a priority","argument":"Anthropic’s post, published September 5, 2025, argues that evaluating and safeguarding against AI-related biological risk is a critical part of responsible AI development. Anthropic says AI could give malicious actors real-time assistance, while biological attacks could have unusually severe consequences and biotechnology advances may lower material barriers. In a controlled trial, participants with Claude developed stronger plans for a hypothetical bioweapons acquisition process than an internet-only group. Anthropic cautions that text-based trials are imperfect proxies for real-world scenarios. Its small wet lab pilot found no evidence of uplift, but Anthropic says it did not provide strong evidence either way; participants in both groups did well. Anthropic is co-sponsoring a larger study to examine whether AI can help non-experts carry out expert-level laboratory tasks. Its conclusion is that uncertainty remains, but the evidence and gravity of biological threats warrant attention from AI developers and policymakers. The larger study’s results and how model capabilities translate into real-world risk remain unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 28, 2026; originally published September 5, 2025. Anthropic says text-based trials are imperfect proxies for real-world scenarios; its wet lab pilot was small and used fairly basic protocols. The larger laboratory study's results are not included in the text.","speaker":{"name":"Anthropic","roleAtSource":"Publisher"},"source":{"title":"LLMs and biorisk","publisher":"Anthropic","format":"Article","url":"https://www.anthropic.com/research/biorisk","publishedAt":"2025-09-05T00:00:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Anthropic argues that evaluating and safeguarding against AI-related biorisk is a critical element of responsible AI development.","excerpt":"we want to establish why we believe that evaluating these risks and safeguarding against them is a critical element of responsible AI development.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says biological attacks are especially concerning because their consequences may be severe and biotechnology advances may reduce material barriers.","excerpt":"the potential consequences of a successful biological attack are unusually severe.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"In Anthropic's controlled trial, participants with Claude 4 models, especially Claude Opus 4, scored higher and had fewer critical failures than the internet-only control group.","excerpt":"Participants with access to Claude 4 models—especially Claude Opus 4—received much higher scores and developed plans with substantially fewer critical failures compared to the internet-only control group.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says text-based uplift trials are imperfect proxies for real-world scenarios, which involve additional factors.","excerpt":"Text-based uplift trials like this are imperfect proxies for real-world scenarios—which involve additional factors like tacit knowledge, materials access, and actor persistence","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic reports that its small wet lab pilot found no evidence of uplift, but says the study did not provide strong evidence either way.","excerpt":"We did not observe any evidence of uplift in this study.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says it is co-sponsoring a larger study to investigate AI assistance on real laboratory tasks, including whether AI can uplift non-experts to expert-level tasks.","excerpt":"this wet lab study will gather evidence by simulating many of the scenarios with which we are concerned, including whether AI can uplift non-experts to the point of being able to carry out expert-level laboratory biology tasks.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says uncertainty remains about threat actors, model capabilities, and how those capabilities translate to risk.","excerpt":"The topic of AI and biorisk is rife with uncertainty—about threat actors, the capabilities of models, and how exactly those capabilities translate to risk.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/biorisk","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What will the larger laboratory study find about whether AI helps non-experts carry out expert-level laboratory biology tasks?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Anthropic argues for treating AI biorisk as a priority","summary":"Anthropic’s post, published September 5, 2025, argues that evaluating and safeguarding against AI-related biological risk is a critical part of responsible AI development. Anthropic says AI could give malicious actors real-time assistance, while biological attacks could have unusually severe consequences and biotechnology advances may lower material barriers. In a controlled trial, participants with Claude developed stronger plans for a hypothetical bioweapons acquisition process than an internet-only group. Anthropic cautions that text-based trials are imperfect proxies for real-world scenarios. Its small wet lab pilot found no evidence of uplift, but Anthropic says it did not provide strong evidence either way; participants in both groups did well. Anthropic is co-sponsoring a larger study to examine whether AI can help non-experts carry out expert-level laboratory tasks. Its conclusion is that uncertainty remains, but the evidence and gravity of biological threats warrant attention from AI developers and policymakers. The larger study’s results and how model capabilities translate into real-world risk remain unknown.","question":"What will the larger laboratory study find about whether AI helps non-experts carry out expert-level laboratory biology tasks?","claims":[{"text":"Anthropic argues that evaluating and safeguarding against AI-related biorisk is a critical element of responsible AI development.","excerpt":"we want to establish why we believe that evaluating these risks and safeguarding against them is a critical element of responsible AI development.","support":"source-stated"},{"text":"Anthropic says biological attacks are especially concerning because their consequences may be severe and biotechnology advances may reduce material barriers.","excerpt":"the potential consequences of a successful biological attack are unusually severe.","support":"source-stated"},{"text":"In Anthropic's controlled trial, participants with Claude 4 models, especially Claude Opus 4, scored higher and had fewer critical failures than the internet-only control group.","excerpt":"Participants with access to Claude 4 models—especially Claude Opus 4—received much higher scores and developed plans with substantially fewer critical failures compared to the internet-only control group.","support":"source-stated"},{"text":"Anthropic says text-based uplift trials are imperfect proxies for real-world scenarios, which involve additional factors.","excerpt":"Text-based uplift trials like this are imperfect proxies for real-world scenarios—which involve additional factors like tacit knowledge, materials access, and actor persistence","support":"source-stated"},{"text":"Anthropic reports that its small wet lab pilot found no evidence of uplift, but says the study did not provide strong evidence either way.","excerpt":"We did not observe any evidence of uplift in this study.","support":"source-stated"},{"text":"Anthropic says it is co-sponsoring a larger study to investigate AI assistance on real laboratory tasks, including whether AI can uplift non-experts to expert-level tasks.","excerpt":"this wet lab study will gather evidence by simulating many of the scenarios with which we are concerned, including whether AI can uplift non-experts to the point of being able to carry out expert-level laboratory biology tasks.","support":"source-stated"},{"text":"Anthropic says uncertainty remains about threat actors, model capabilities, and how those capabilities translate to risk.","excerpt":"The topic of AI and biorisk is rife with uncertainty—about threat actors, the capabilities of models, and how exactly those capabilities translate to risk.","support":"source-stated"}],"limitations":["Text as retrieved on October 28, 2026; originally published September 5, 2025.","Anthropic says text-based trials are imperfect proxies for real-world scenarios; its wet lab pilot was small and used fairly basic protocols.","The larger laboratory study's results are not included in the text."],"mediaId":null,"correctionReason":null,"id":"1594a8990a7e1d707e9ca08b0aaa1dc05e01ac6eb6c6b4a6849ee2738ef14e8e","publicationId":"eb4f9a8d-0de2-1f2d-0183-c227ec88c102","version":1,"sourceUrl":"https://www.anthropic.com/research/biorisk","sourceTitle":"LLMs and biorisk","publisher":"Anthropic","publishedAt":"2025-09-05T00:00:00.000Z","originalPublished":{"value":"2025-09-05T00:00:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-20T10:51:21.988Z","createdAt":"2026-11-12T07:18:23.717Z","historical":true,"entityIds":["anthropic"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/biorisk","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"cf3777e8846df90793469b80331cbc19050813c58d1ca22739429cc2ff0596b9","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-28T04:34:19.133Z"},"provenance":{"textHash":"a7fcaf6f75937fd6920ad43df052b3fadd63cb37a415b2c953fbb2b1dc28c39e","observationId":"f34fb381390198f388df830c616cc6745cd9b1d92f5024410d5cd2bd12291039","retrievedAt":"2026-10-28T04:35:48.940Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"90c96346-9776-ca20-2a33-f025a92d6a3f"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"f3ef3798-64ab-47fd-791c-3600ea244165"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-12T07:13:59.281Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/anthropic-argues-for-treating-ai-biorisk-as-a-priority-1594a899","formats":{"html":"/readings/anthropic-argues-for-treating-ai-biorisk-as-a-priority-1594a899","markdown":"/readings/anthropic-argues-for-treating-ai-biorisk-as-a-priority-1594a899.md","json":"/readings/anthropic-argues-for-treating-ai-biorisk-as-a-priority-1594a899.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[],"entities":[{"id":"anthropic","name":"Anthropic","kind":"organization","context":"Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.","referenceUrls":["https://www.anthropic.com/news","https://www.anthropic.com/research/team/economics"],"path":"/entities/anthropic","json":"/entities/anthropic.json","markdown":"/entities/anthropic.md"},{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.md"},{"id":"claude","name":"Claude","kind":"model","context":"Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.","referenceUrls":["https://claude.com/product/overview","https://www.anthropic.com/news","https://www.anthropic.com/research/team/economics"],"path":"/entities/claude","json":"/entities/claude.json","markdown":"/entities/claude.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"1594a8990a7e1d707e9ca08b0aaa1dc05e01ac6eb6c6b4a6849ee2738ef14e8e","revision":"eb4f9a8d-0de2-1f2d-0183-c227ec88c102","revisionKind":"publication-id","title":"Anthropic argues for treating AI biorisk as a priority","url":"https://escapemeridian.com/readings/anthropic-argues-for-treating-ai-biorisk-as-a-priority-1594a899"},"source":{"type":"source","id":"https://www.anthropic.com/research/biorisk","revision":"a7fcaf6f75937fd6920ad43df052b3fadd63cb37a415b2c953fbb2b1dc28c39e","revisionKind":"sha256-source-text","title":"LLMs and biorisk","url":"https://www.anthropic.com/research/biorisk"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"1594a8990a7e1d707e9ca08b0aaa1dc05e01ac6eb6c6b4a6849ee2738ef14e8e","title":"Anthropic argues for treating AI biorisk as a priority","argument":"Anthropic’s post, published September 5, 2025, argues that evaluating and safeguarding against AI-related biological risk is a critical part of responsible AI development. Anthropic says AI could give malicious actors real-time assistance, while biological attacks could have unusually severe consequences and biotechnology advances may lower material barriers. In a controlled trial, participants with Claude developed stronger plans for a hypothetical bioweapons acquisition process than an internet-only group. Anthropic cautions that text-based trials are imperfect proxies for real-world scenarios. Its small wet lab pilot found no evidence of uplift, but Anthropic says it did not provide strong evidence either way; participants in both groups did well. Anthropic is co-sponsoring a larger study to examine whether AI can help non-experts carry out expert-level laboratory tasks. Its conclusion is that uncertainty remains, but the evidence and gravity of biological threats warrant attention from AI developers and policymakers. The larger study’s results and how model capabilities translate into real-world risk remain unknown.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 28, 2026; originally published September 5, 2025. Anthropic says text-based trials are imperfect proxies for real-world scenarios; its wet lab pilot was small and used fairly basic protocols. The larger laboratory study's results are not included in the text.","speaker":{"name":"Anthropic","roleAtSource":"Publisher"},"source":{"title":"LLMs and biorisk","publisher":"Anthropic","format":"Article","url":"https://www.anthropic.com/research/biorisk","publishedAt":"2025-09-05T00:00:00.000Z","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Anthropic argues that evaluating and safeguarding against AI-related biorisk is a critical element of responsible AI development.","excerpt":"we want to establish why we believe that evaluating these risks and safeguarding against them is a critical element of responsible AI development.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says biological attacks are especially concerning because their consequences may be severe and biotechnology advances may reduce material barriers.","excerpt":"the potential consequences of a successful biological attack are unusually severe.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"In Anthropic's controlled trial, participants with Claude 4 models, especially Claude Opus 4, scored higher and had fewer critical failures than the internet-only control group.","excerpt":"Participants with access to Claude 4 models—especially Claude Opus 4—received much higher scores and developed plans with substantially fewer critical failures compared to the internet-only control group.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says text-based uplift trials are imperfect proxies for real-world scenarios, which involve additional factors.","excerpt":"Text-based uplift trials like this are imperfect proxies for real-world scenarios—which involve additional factors like tacit knowledge, materials access, and actor persistence","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic reports that its small wet lab pilot found no evidence of uplift, but says the study did not provide strong evidence either way.","excerpt":"We did not observe any evidence of uplift in this study.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says it is co-sponsoring a larger study to investigate AI assistance on real laboratory tasks, including whether AI can uplift non-experts to expert-level tasks.","excerpt":"this wet lab study will gather evidence by simulating many of the scenarios with which we are concerned, including whether AI can uplift non-experts to the point of being able to carry out expert-level laboratory biology tasks.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"},{"text":"Anthropic says uncertainty remains about threat actors, model capabilities, and how those capabilities translate to risk.","excerpt":"The topic of AI and biorisk is rife with uncertainty—about threat actors, the capabilities of models, and how exactly those capabilities translate to risk.","url":"https://www.anthropic.com/research/biorisk","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/biorisk","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What will the larger laboratory study find about whether AI helps non-experts carry out expert-level laboratory biology tasks?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Anthropic argues for treating AI biorisk as a priority","summary":"Anthropic’s post, published September 5, 2025, argues that evaluating and safeguarding against AI-related biological risk is a critical part of responsible AI development. Anthropic says AI could give malicious actors real-time assistance, while biological attacks could have unusually severe consequences and biotechnology advances may lower material barriers. In a controlled trial, participants with Claude developed stronger plans for a hypothetical bioweapons acquisition process than an internet-only group. Anthropic cautions that text-based trials are imperfect proxies for real-world scenarios. Its small wet lab pilot found no evidence of uplift, but Anthropic says it did not provide strong evidence either way; participants in both groups did well. Anthropic is co-sponsoring a larger study to examine whether AI can help non-experts carry out expert-level laboratory tasks. Its conclusion is that uncertainty remains, but the evidence and gravity of biological threats warrant attention from AI developers and policymakers. The larger study’s results and how model capabilities translate into real-world risk remain unknown.","question":"What will the larger laboratory study find about whether AI helps non-experts carry out expert-level laboratory biology tasks?","claims":[{"text":"Anthropic argues that evaluating and safeguarding against AI-related biorisk is a critical element of responsible AI development.","excerpt":"we want to establish why we believe that evaluating these risks and safeguarding against them is a critical element of responsible AI development.","support":"source-stated"},{"text":"Anthropic says biological attacks are especially concerning because their consequences may be severe and biotechnology advances may reduce material barriers.","excerpt":"the potential consequences of a successful biological attack are unusually severe.","support":"source-stated"},{"text":"In Anthropic's controlled trial, participants with Claude 4 models, especially Claude Opus 4, scored higher and had fewer critical failures than the internet-only control group.","excerpt":"Participants with access to Claude 4 models—especially Claude Opus 4—received much higher scores and developed plans with substantially fewer critical failures compared to the internet-only control group.","support":"source-stated"},{"text":"Anthropic says text-based uplift trials are imperfect proxies for real-world scenarios, which involve additional factors.","excerpt":"Text-based uplift trials like this are imperfect proxies for real-world scenarios—which involve additional factors like tacit knowledge, materials access, and actor persistence","support":"source-stated"},{"text":"Anthropic reports that its small wet lab pilot found no evidence of uplift, but says the study did not provide strong evidence either way.","excerpt":"We did not observe any evidence of uplift in this study.","support":"source-stated"},{"text":"Anthropic says it is co-sponsoring a larger study to investigate AI assistance on real laboratory tasks, including whether AI can uplift non-experts to expert-level tasks.","excerpt":"this wet lab study will gather evidence by simulating many of the scenarios with which we are concerned, including whether AI can uplift non-experts to the point of being able to carry out expert-level laboratory biology tasks.","support":"source-stated"},{"text":"Anthropic says uncertainty remains about threat actors, model capabilities, and how those capabilities translate to risk.","excerpt":"The topic of AI and biorisk is rife with uncertainty—about threat actors, the capabilities of models, and how exactly those capabilities translate to risk.","support":"source-stated"}],"limitations":["Text as retrieved on October 28, 2026; originally published September 5, 2025.","Anthropic says text-based trials are imperfect proxies for real-world scenarios; its wet lab pilot was small and used fairly basic protocols.","The larger laboratory study's results are not included in the text."],"mediaId":null,"correctionReason":null,"id":"1594a8990a7e1d707e9ca08b0aaa1dc05e01ac6eb6c6b4a6849ee2738ef14e8e","publicationId":"eb4f9a8d-0de2-1f2d-0183-c227ec88c102","version":1,"sourceUrl":"https://www.anthropic.com/research/biorisk","sourceTitle":"LLMs and biorisk","publisher":"Anthropic","publishedAt":"2025-09-05T00:00:00.000Z","originalPublished":{"value":"2025-09-05T00:00:00.000Z","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-20T10:51:21.988Z","createdAt":"2026-11-12T07:18:23.717Z","historical":true,"entityIds":["anthropic"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.anthropic.com/research/biorisk","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"cf3777e8846df90793469b80331cbc19050813c58d1ca22739429cc2ff0596b9","rubricVersion":"unrecorded","provider":"anthropic","model":"claude-opus-5-5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-28T04:34:19.133Z"},"provenance":{"textHash":"a7fcaf6f75937fd6920ad43df052b3fadd63cb37a415b2c953fbb2b1dc28c39e","observationId":"f34fb381390198f388df830c616cc6745cd9b1d92f5024410d5cd2bd12291039","retrievedAt":"2026-10-28T04:35:48.940Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"90c96346-9776-ca20-2a33-f025a92d6a3f"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"f3ef3798-64ab-47fd-791c-3600ea244165"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-12T07:13:59.281Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"eb4f9a8d-0de2-1f2d-0183-c227ec88c102","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading consistently attributes the arguments, trial results and plans to Anthropic's post. It does not present Anthropic's assertions as established fact or as Escape Meridian's view.","excerpt":"In this post, we want to expand on our perspective on AI and biological risk (biorisk)."},{"dimension":"summary","status":"uncertain","explanation":"The central term \"uplift\" is used without explanation, and no definitions are supplied. A general reader may not know it means AI measurably improving someone's ability compared with internet access alone. \"Wet lab\" is also unexplained. The reading also drops a key trial condition: the Claude group used Claude \"with safeguards removed.\" Smallest repair: briefly gloss uplift and wet lab, and add that the trial used an unsafeguarded Claude. The title, question and conclusion otherwise preserve the post's argument and its uncertainty.","excerpt":"the model-assisted group had additional access to Claude with safeguards removed."},{"dimension":"date","status":"pass","explanation":"The stated date, September 5, 2025, matches the UTC day of evidence.originalPublished. The limitations note the later retrieval date. Nothing is presented with hindsight. The larger study is reported as planned and its results as unknown. The page was modified on 2026-10-23, but the cited passages are integral to the original article. Its references to Claude Opus 4 and the US AI Action Plan are consistent with a September 2025 date.","excerpt":"We look forward to learning more from the results of this study, which will be important to informing our assessment of risk."},{"dimension":"support","status":"pass","explanation":"Each claim matches its excerpt, and the claims are worded as Anthropic's statements. The account of the wet lab pilot keeps the caveat that it gave no strong evidence either way. The caution that text trials are imperfect proxies is also preserved. Edited-after-publication status was considered: the cited text is plainly part of the original argument. The question follows from the planned study without assuming its outcome.","excerpt":"So although this pilot study did not provide strong evidence one way or another about AI uplift, it did suggest that tacit knowledge may not be as significant of a bottleneck as many have suggested."},{"dimension":"media","status":"pass","explanation":"No media is selected and no verified media candidates were supplied, so nothing about media identity or relevance needs checking.","excerpt":null}],"createdAt":"2026-11-12T07:13:59.281Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}