{"id":"427658ecb9e9bc9b5cf69e79fb99611c5dd087bb7ee3f45efd73d708891fcf46","title":"Google DeepMind introduces evaluations of dangerous AI capabilities","argument":"On March 21, 2024, Google DeepMind introduces a programme of evaluations for dangerous capabilities in Gemini models. The evaluations cover persuasion and deception, cyber-security, self-proliferation, self-reasoning and self-modification, and biological and nuclear risk. Google DeepMind reports no evidence of strong dangerous capabilities in the models it evaluated, but flags early warning signs. It says the programme is intended to lay the groundwork for rigorous evaluation of dangerous capabilities as more capable models are developed. The abstract does not specify what the early warning signs were or how the evaluations performed across the five topics.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published March 21, 2024. The abstract does not describe the early warning signs or provide results for individual evaluation topics.","speaker":{"name":"Google DeepMind","roleAtSource":"Publisher"},"source":{"title":"Evaluating Frontier Models for Dangerous Capabilities","publisher":"Google DeepMind","format":"Article","url":"https://deepmind.google/research/publications/78150/","publishedAt":"2024-03-21T00:00:00+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Google DeepMind introduces a programme of dangerous-capability evaluations and pilots them on Gemini models.","excerpt":"we introduce a programme of new \"dangerous capability\" evaluations and pilot them on Gemini models.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"},{"text":"The evaluations span five topics: persuasion and deception; cyber-security; self-proliferation; self-reasoning and self-modification; and biological and nuclear risk.","excerpt":"These evaluations cover five topics: (1) persuasion & deception; (2) cyber-security; (3) self-proliferation; (4) self-reasoning & self-modification; and (5) biological and nuclear risk.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"},{"text":"Google DeepMind reports no evidence of strong dangerous capabilities in the evaluated models, while flagging early warning signs.","excerpt":"We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"},{"text":"Google DeepMind says the programme aims to lay groundwork for rigorous evaluation in preparation for future, more capable models.","excerpt":"Our goal is to help lay the groundwork for a rigorous science of dangerous capability evaluation, in preparation for future, more capable models.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/78150/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What were the early warning signs, and how did they vary across the evaluation topics?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Google DeepMind introduces evaluations of dangerous AI capabilities","summary":"On March 21, 2024, Google DeepMind introduces a programme of evaluations for dangerous capabilities in Gemini models. The evaluations cover persuasion and deception, cyber-security, self-proliferation, self-reasoning and self-modification, and biological and nuclear risk. Google DeepMind reports no evidence of strong dangerous capabilities in the models it evaluated, but flags early warning signs. It says the programme is intended to lay the groundwork for rigorous evaluation of dangerous capabilities as more capable models are developed. The abstract does not specify what the early warning signs were or how the evaluations performed across the five topics.","question":"What were the early warning signs, and how did they vary across the evaluation topics?","claims":[{"text":"Google DeepMind introduces a programme of dangerous-capability evaluations and pilots them on Gemini models.","excerpt":"we introduce a programme of new \"dangerous capability\" evaluations and pilot them on Gemini models.","support":"source-stated"},{"text":"The evaluations span five topics: persuasion and deception; cyber-security; self-proliferation; self-reasoning and self-modification; and biological and nuclear risk.","excerpt":"These evaluations cover five topics: (1) persuasion & deception; (2) cyber-security; (3) self-proliferation; (4) self-reasoning & self-modification; and (5) biological and nuclear risk.","support":"source-stated"},{"text":"Google DeepMind reports no evidence of strong dangerous capabilities in the evaluated models, while flagging early warning signs.","excerpt":"We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs.","support":"source-stated"},{"text":"Google DeepMind says the programme aims to lay groundwork for rigorous evaluation in preparation for future, more capable models.","excerpt":"Our goal is to help lay the groundwork for a rigorous science of dangerous capability evaluation, in preparation for future, more capable models.","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published March 21, 2024.","The abstract does not describe the early warning signs or provide results for individual evaluation topics."],"mediaId":null,"correctionReason":null,"id":"427658ecb9e9bc9b5cf69e79fb99611c5dd087bb7ee3f45efd73d708891fcf46","publicationId":"e3aaa9f1-a7e4-0a1a-c0ec-78fcf2287e9b","version":1,"sourceUrl":"https://deepmind.google/research/publications/78150/","sourceTitle":"Evaluating Frontier Models for Dangerous Capabilities","publisher":"Google DeepMind","publishedAt":"2024-03-21T00:00:00+00:00","originalPublished":{"value":"2024-03-21T00:00:00+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-27T23:21:58.932Z","createdAt":"2026-11-20T19:18:11.213Z","historical":true,"entityIds":["google-deepmind"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/78150/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"ebf2611a2bf5c28ec71a8e8500a64505fb36056c14887c83e549d57a65f19c5c","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.52,"decidedAt":"2026-10-27T23:21:57.392Z"},"provenance":{"textHash":"40a04f3dfc91bd3923a2c6f7831187fc29d9cfdac8789ebf65e5c4dfa2b858b7","observationId":"ac9aa4c96da14e987f90eb67f069edd475cb673aba174e5ba4031c645ec318e3","retrievedAt":"2026-10-27T23:23:52.485Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"892ae49a-fe8d-b249-d3da-d8b039a64751"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"a4344b74-b7ab-23e8-1511-0b2ff8cf4d56"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-20T19:13:27.733Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/google-deepmind-introduces-evaluations-of-dangerous-ai-427658ec","formats":{"html":"/readings/google-deepmind-introduces-evaluations-of-dangerous-ai-427658ec","markdown":"/readings/google-deepmind-introduces-evaluations-of-dangerous-ai-427658ec.md","json":"/readings/google-deepmind-introduces-evaluations-of-dangerous-ai-427658ec.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":[],"conditions":[],"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":"google-deepmind","name":"Google DeepMind","kind":"organization","context":"Google DeepMind develops AI systems for general tasks and scientific research, including AlphaFold.","referenceUrls":["https://deepmind.google/blog/","https://deepmind.google/science/alphafold/"],"path":"/entities/google-deepmind","json":"/entities/google-deepmind.json","markdown":"/entities/google-deepmind.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"427658ecb9e9bc9b5cf69e79fb99611c5dd087bb7ee3f45efd73d708891fcf46","revision":"e3aaa9f1-a7e4-0a1a-c0ec-78fcf2287e9b","revisionKind":"publication-id","title":"Google DeepMind introduces evaluations of dangerous AI capabilities","url":"https://escapemeridian.com/readings/google-deepmind-introduces-evaluations-of-dangerous-ai-427658ec"},"source":{"type":"source","id":"https://deepmind.google/research/publications/78150/","revision":"40a04f3dfc91bd3923a2c6f7831187fc29d9cfdac8789ebf65e5c4dfa2b858b7","revisionKind":"sha256-source-text","title":"Evaluating Frontier Models for Dangerous Capabilities","url":"https://deepmind.google/research/publications/78150/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"427658ecb9e9bc9b5cf69e79fb99611c5dd087bb7ee3f45efd73d708891fcf46","title":"Google DeepMind introduces evaluations of dangerous AI capabilities","argument":"On March 21, 2024, Google DeepMind introduces a programme of evaluations for dangerous capabilities in Gemini models. The evaluations cover persuasion and deception, cyber-security, self-proliferation, self-reasoning and self-modification, and biological and nuclear risk. Google DeepMind reports no evidence of strong dangerous capabilities in the models it evaluated, but flags early warning signs. It says the programme is intended to lay the groundwork for rigorous evaluation of dangerous capabilities as more capable models are developed. The abstract does not specify what the early warning signs were or how the evaluations performed across the five topics.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 27, 2026; originally published March 21, 2024. The abstract does not describe the early warning signs or provide results for individual evaluation topics.","speaker":{"name":"Google DeepMind","roleAtSource":"Publisher"},"source":{"title":"Evaluating Frontier Models for Dangerous Capabilities","publisher":"Google DeepMind","format":"Article","url":"https://deepmind.google/research/publications/78150/","publishedAt":"2024-03-21T00:00:00+00:00","dateBasis":"Date stated on the publisher’s page."},"evidence":[{"text":"Google DeepMind introduces a programme of dangerous-capability evaluations and pilots them on Gemini models.","excerpt":"we introduce a programme of new \"dangerous capability\" evaluations and pilot them on Gemini models.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"},{"text":"The evaluations span five topics: persuasion and deception; cyber-security; self-proliferation; self-reasoning and self-modification; and biological and nuclear risk.","excerpt":"These evaluations cover five topics: (1) persuasion & deception; (2) cyber-security; (3) self-proliferation; (4) self-reasoning & self-modification; and (5) biological and nuclear risk.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"},{"text":"Google DeepMind reports no evidence of strong dangerous capabilities in the evaluated models, while flagging early warning signs.","excerpt":"We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"},{"text":"Google DeepMind says the programme aims to lay groundwork for rigorous evaluation in preparation for future, more capable models.","excerpt":"Our goal is to help lay the groundwork for a rigorous science of dangerous capability evaluation, in preparation for future, more capable models.","url":"https://deepmind.google/research/publications/78150/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/78150/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What were the early warning signs, and how did they vary across the evaluation topics?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Google DeepMind introduces evaluations of dangerous AI capabilities","summary":"On March 21, 2024, Google DeepMind introduces a programme of evaluations for dangerous capabilities in Gemini models. The evaluations cover persuasion and deception, cyber-security, self-proliferation, self-reasoning and self-modification, and biological and nuclear risk. Google DeepMind reports no evidence of strong dangerous capabilities in the models it evaluated, but flags early warning signs. It says the programme is intended to lay the groundwork for rigorous evaluation of dangerous capabilities as more capable models are developed. The abstract does not specify what the early warning signs were or how the evaluations performed across the five topics.","question":"What were the early warning signs, and how did they vary across the evaluation topics?","claims":[{"text":"Google DeepMind introduces a programme of dangerous-capability evaluations and pilots them on Gemini models.","excerpt":"we introduce a programme of new \"dangerous capability\" evaluations and pilot them on Gemini models.","support":"source-stated"},{"text":"The evaluations span five topics: persuasion and deception; cyber-security; self-proliferation; self-reasoning and self-modification; and biological and nuclear risk.","excerpt":"These evaluations cover five topics: (1) persuasion & deception; (2) cyber-security; (3) self-proliferation; (4) self-reasoning & self-modification; and (5) biological and nuclear risk.","support":"source-stated"},{"text":"Google DeepMind reports no evidence of strong dangerous capabilities in the evaluated models, while flagging early warning signs.","excerpt":"We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs.","support":"source-stated"},{"text":"Google DeepMind says the programme aims to lay groundwork for rigorous evaluation in preparation for future, more capable models.","excerpt":"Our goal is to help lay the groundwork for a rigorous science of dangerous capability evaluation, in preparation for future, more capable models.","support":"source-stated"}],"limitations":["Text as retrieved on October 27, 2026; originally published March 21, 2024.","The abstract does not describe the early warning signs or provide results for individual evaluation topics."],"mediaId":null,"correctionReason":null,"id":"427658ecb9e9bc9b5cf69e79fb99611c5dd087bb7ee3f45efd73d708891fcf46","publicationId":"e3aaa9f1-a7e4-0a1a-c0ec-78fcf2287e9b","version":1,"sourceUrl":"https://deepmind.google/research/publications/78150/","sourceTitle":"Evaluating Frontier Models for Dangerous Capabilities","publisher":"Google DeepMind","publishedAt":"2024-03-21T00:00:00+00:00","originalPublished":{"value":"2024-03-21T00:00:00+00:00","precision":"instant","basis":"backfill:article_published"},"firstObservedAt":"2026-10-27T23:21:58.932Z","createdAt":"2026-11-20T19:18:11.213Z","historical":true,"entityIds":["google-deepmind"],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://deepmind.google/research/publications/78150/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"ebf2611a2bf5c28ec71a8e8500a64505fb36056c14887c83e549d57a65f19c5c","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.52,"decidedAt":"2026-10-27T23:21:57.392Z"},"provenance":{"textHash":"40a04f3dfc91bd3923a2c6f7831187fc29d9cfdac8789ebf65e5c4dfa2b858b7","observationId":"ac9aa4c96da14e987f90eb67f069edd475cb673aba174e5ba4031c645ec318e3","retrievedAt":"2026-10-27T23:23:52.485Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"892ae49a-fe8d-b249-d3da-d8b039a64751"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"a4344b74-b7ab-23e8-1511-0b2ff8cf4d56"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-20T19:13:27.733Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"e3aaa9f1-a7e4-0a1a-c0ec-78fcf2287e9b","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading attributes the evaluation programme, findings and goal to Google DeepMind. That fits this publisher page for a paper by DeepMind-affiliated authors. Results are kept as the source's own reports, not as established facts.","excerpt":"we introduce a programme of new \"dangerous capability\" evaluations and pilot them on Gemini models."},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful to the abstract and names the gap about the warning signs. But central terms are left unexplained for a non-specialist, and the definitions list is empty: \"self-proliferation\", \"self-reasoning and self-modification\" and \"dangerous capability evaluations\". Readers may not see what is being tested. Smallest repair: add term-library tooltips for these terms. Also, \"rigorous evaluation\" drops the source's \"rigorous science of\" framing.","excerpt":"(3) self-proliferation; (4) self-reasoning & self-modification"},{"dimension":"date","status":"pass","explanation":"The page dateline reads March 21, 2024, matching evidence.originalPublished stated at day precision. The page was modified in 2025, but the reading relies only on the abstract, which is plainly the original piece. Nothing after the publication date or in hindsight is presented. This is a publisher page, not an archive copy, so no archive disclosure is needed.","excerpt":"March 21, 2024\nEvaluating Frontier Models for Dangerous Capabilities"},{"dimension":"support","status":"pass","explanation":"Each claim matches the abstract verbatim in substance. The findings are kept as reported: no strong dangerous capabilities, but early warning signs. The stated goal stays attributed. The reading claims no benefits or real-world reliability beyond the source. The question follows from an unresolved gap in the abstract.","excerpt":"We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs."},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified media candidates were supplied, so there is no media identity or relevance issue.","excerpt":null}],"createdAt":"2026-11-20T19:13:27.733Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}