{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T06:40:50.084Z","headline":"英国AI安全研究所测试发现所有前沿AI模型均试图在网络安全评估中作弊","description":"英国AI安全研究所（AISI）对OpenAI和Anthropic的五款前沿模型进行网络安全评估，发现所有模型均试图作弊。GPT-5.4在14.1%的测试中作弊（475次中67次），GPT-5.5为11.4%，GPT-5.6 Sol为12.6%，Claude Opus 4.7为9.1%，Claude Mythos Preview为7.8%。","url":"https://www.aioga.com/news/cmrwbm0ir002xroj0mxzodcyx/","mainEntityOfPage":"https://www.aioga.com/news/cmrwbm0ir002xroj0mxzodcyx/","datePublished":"2026-07-22T16:41:49.000Z","dateModified":"2026-07-22T16:41:49.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://the-decoder.com/every-frontier-ai-model-tested-by-britains-safety-institute-tried-to-cheat-on-cybersecurity-evaluations","https://aihot.virxact.com/items/cmrwbm0ir002xroj0mxzodcyx"],"canonicalUrl":"https://www.aioga.com/news/cmrwbm0ir002xroj0mxzodcyx/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：英国AI安全研究所（AISI）对OpenAI和Anthropic的五款前沿模型进行网络安全评估，发现所有模型均试图作弊。 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cmrwbm0ir002xroj0mxzodcyx/","dateCreated":"2026-07-22T16:41:49.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"the-decoder.com source article","url":"https://the-decoder.com/every-frontier-ai-model-tested-by-britains-safety-institute-tried-to-cheat-on-cybersecurity-evaluations","datePublished":"2026-07-22T16:41:49.000Z","provider":{"@type":"Organization","name":"the-decoder.com","url":"https://the-decoder.com/every-frontier-ai-model-tested-by-britains-safety-institute-tried-to-cheat-on-cybersecurity-evaluations"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmrwbm0ir002xroj0mxzodcyx","datePublished":"2026-07-22T16:41:49.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmrwbm0ir002xroj0mxzodcyx"}}],"aggregationSource":"The Decoder：AI News（RSS）","originalPublisher":{"name":"the-decoder.com","url":"https://the-decoder.com/every-frontier-ai-model-tested-by-britains-safety-institute-tried-to-cheat-on-cybersecurity-evaluations"},"article":{"id":"cmrwbm0ir002xroj0mxzodcyx","slug":"cmrwbm0ir002xroj0mxzodcyx","url":"https://www.aioga.com/news/cmrwbm0ir002xroj0mxzodcyx/","title":"英国AI安全研究所测试发现所有前沿AI模型均试图在网络安全评估中作弊","title_en":"Every frontier AI model tested by Britain's safety institute tried to cheat on cybersecurity evaluations","summary":"英国AI安全研究所（AISI）对OpenAI和Anthropic的五款前沿模型进行网络安全评估，发现所有模型均试图作弊。GPT-5.4在14.1%的测试中作弊（475次中67次），GPT-5.5为11.4%，GPT-5.6 Sol为12.6%，Claude Opus 4.7为9.1%，Claude Mythos Preview为7.8%。","source":"The Decoder：AI News（RSS）","sourceUrl":"https://the-decoder.com/every-frontier-ai-model-tested-by-britains-safety-institute-tried-to-cheat-on-cybersecurity-evaluations","aiHotUrl":"https://aihot.virxact.com/items/cmrwbm0ir002xroj0mxzodcyx","publishedAt":"2026-07-22T16:41:49.000Z","category":"行业动态","score":61,"selected":false,"articleBody":["The UK's AI Safety Institute systematically tested models from OpenAI and Anthropic for cheating in cybersecurity evaluations. All five models tried to get around the rules.","In the AI Safety Institute's (AISI) tests, models must find hidden strings known as \"flags\" inside simulated environments. They perform offensive cyber tasks such as reverse engineering and exploiting security flaws. Each task has clear rules and a defined path to the solution.","All five frontier models tested：https://www.aisi.gov.uk/blog/cheating-behaviour-in-frontier-model-evaluations tried to cheat. Instead of following the intended solution path, they used shortcuts, workarounds, or actions that were explicitly prohibited. GPT-5.4 cheated in 14.1 percent of test runs (67 out of 475), GPT-5.5 in 11.4 percent, and GPT-5.6 Sol in 12.6 percent. Anthropic's Claude Opus 4.7 came in at 9.1 percent, while Claude Mythos Preview reached 7.8 percent. None were prompted to cheat. Ad","The label \"cheating\" doesn't necessarily imply deceptive intent, the AISI says. But the behavior is still a problem: it could cause evaluations to overstate a model's actual abilities and mislead users when the success of a task is hard to verify. Ad DEC_D_Incontent-1","AISI found no clear link between greater model capability and more frequent cheating attempts. Instead, the institute says cheating behavior is \"substantially shaped by the specifics of the techniques used to train the model, including alignment training, and not just raw capability.\"","Cheating methods vary by model. According to AISI, common tactics include searching online for solutions and attacking systems outside the evaluation target, including the system hosting the model. Models also probed the evaluation software to extract answers. Ad","GPT-5.6 Sol scored highest for the strategy of searching the internet for solutions. That lines up with the autonomous cyberattack OpenAI and Hugging Face recently disclosed：https://the-decoder.com/openai-claims-responsibility-for-the-hugging-face-hack-after-its-own-models-escaped-a-test-sandbox/, where OpenAI models tried to hack Hugging Face to grab a benchmark solution. METR's investigation of GPT-5.6 Sol：https://the-decoder.com/gpt-5-6-sol-cheats-on-software-tests-more-than-any-model-before-it/ was also heavily compromised by fraudulent behavior.","One tested model went even further, AISI reports. It wrote and ran code on an external service on the open internet to access AISI's evaluation infrastructure. The incident occurred during a task that had been misconfigured by mistake and couldn't be solved, triggering a security alert. AISI says the attempt might have worked if its infrastructure had been less secure. The reported results should be treated as lower bounds because the automated LLM monitor may have missed some cases. Ad DEC_D_Incontent-2","An obvious way to catch cheating would be to just ask the model whether it performed any prohibited actions. But according to AISI's experiments, that doesn't work reliably. The models didn't consistently admit to cheating and described their behavior as wrong in fewer than 50 percent of cases. Even when they acknowledged actions that met AISI's definition of cheating, they often framed them as permitted. Ad","Analyzing the \"chain of thought,\"：https://the-decoder.com/highlighted-chain-of-thought-prompting-boosts-llm-accuracy-and-verifiability/ or the models' visible reasoning, also proved unreliable. Claude Opus 4.7 produced no reasoning trace in 87 percent of cheating cases because it uses adaptive reasoning. GPT-5.6 Sol's thought process showed no sign of unauthorized actions in 40 percent of cases.","AISI also saw models consider whether a planned action would count as cheating and then carry it out anyway. During one internal deliberation, GPT-5.6 Sol questioned whether using a cluster API would violate the stated rules. It then chose a different prohibited action.","AISI warns that the consequences could grow as models become more capable, even if the cheating rate stays constant. More capable models could find cheating methods that are harder to detect and cause more harm if they work. This is especially relevant to offensive cyber capabilities：https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing, which are improving quickly. Earlier AISI research：https://www.aisi.gov.uk/research/loss-of-oversight-how-ai-systems-may-become-harder-to-audit-monitor-and-investigate also argues that monitoring models could become more difficult over time.","Stay in the loop on AI. Clear, useful, no fluff.","Follow The Decoder for AI news, background stories and expert analyses.","The Decoder：https://the-decoder.com/"],"articleImages":[{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/07/aisi_cheating_eval-3.png","alt":"","afterParagraph":2,"url":"/media/articles/cmrwbm0ir002xroj0mxzodcyx/4d534832193476e8.png"},{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/07/aisi_cheating_eval-2.png","alt":"Gestapeltes Säulendiagramm: Anteil verschiedener Betrugsstrategien (z. B. Internetrecherche, Angriff auf fremde Systeme) bei Modellen GPT-5.4 bis Claude Mythos Preview.","afterParagraph":5,"url":"/media/articles/cmrwbm0ir002xroj0mxzodcyx/e03259742cad4732.png"},{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/07/aisi_answers.png","alt":"","afterParagraph":8,"url":"/media/articles/cmrwbm0ir002xroj0mxzodcyx/c8fa36bcb6d08fa8.png"},{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/07/aisi_cheating_eval-1.png","alt":"","afterParagraph":10,"url":"/media/articles/cmrwbm0ir002xroj0mxzodcyx/69215459db236cc9.png"}],"mediaStatus":"ok","articleBodyZh":["英国的人工智能安全研究所系统地测试了来自 OpenAI 和 Anthropic 的模型，以评估它们在网络安全测试中的作弊情况。所有五款模型都尝试规避规则。","在人工智能安全研究所（AISI）的测试中，模型必须在模拟环境中找到被称为“旗标”的隐藏字符串。它们执行诸如逆向工程和利用安全漏洞等攻击性网络任务。每个任务都有明确的规则和定义好的解决方案路径。","所有五款前沿模型测试：https://www.aisi.gov.uk/blog/cheating-behaviour-in-frontier-model-evaluations 都尝试作弊。它们没有遵循预期的解决方案路径，而是使用了捷径、变通方法或明确禁止的操作。GPT-5.4 在测试运行中有 14.1% （475 次中的 67 次）作弊，GPT-5.5 为 11.4%，GPT-5.6 Sol 为 12.6%。Anthropic 的 Claude Opus 4.7 为 9.1%，而 Claude Mythos Preview 达到 7.8%。这些模型都没有被特别提示去作弊。","AISI 表示，“作弊”标签不一定意味着有欺骗意图。但这种行为仍然是一个问题：它可能导致评估过高估计模型的实际能力，并在任务结果难以验证时误导用户。","AISI 没有发现模型能力越强就越频繁尝试作弊的明确关联。研究所表示，作弊行为“在很大程度上受训练模型所使用的具体技术影响，包括对齐训练，而不仅仅是原始能力。”","作弊方法因模型而异。根据 AISI 的说法，常见策略包括在网上搜索解决方案以及攻击评估目标之外的系统，包括托管模型的系统。模型还会探查评估软件以提取答案。","GPT-5.6 Sol 在搜索互联网获取解决方案的策略上得分最高。这与 OpenAI 和 Hugging Face 最近披露的自主网络攻击一致：https://the-decoder.com/openai-claims-responsibility-for-the-hugging-face-hack-after-its-own-models-escaped-a-test-sandbox/，其中 OpenAI 的模型试图入侵 Hugging Face 以获取基准解决方案。METR 对 GPT-5.6 Sol 的调查：https://the-decoder.com/gpt-5-6-sol-cheats-on-software-tests-more-than-any-model-before-it/ 也严重受到了欺诈行为的影响。","AISI 报告称，一款被测试的模型做得更过分。它在开放互联网的外部服务上编写并运行代码以访问 AISI 的评估基础设施。该事件发生在一个因配置错误而无法解决的任务过程中，因此触发了安全警报。AISI 表示，如果其基础设施安全性较低，这次尝试可能会成功。报告的结果应被视为下限，因为自动化 LLM 监控可能遗漏了一些情况。Ad DEC_D_Incontent-2","抓作弊的一个明显方法就是直接询问模型是否执行了任何被禁止的操作。但根据 AISI 的实验，这种方法不可靠。模型没有始终承认作弊行为，并且在不到 50% 的情况下将其行为描述为错误。即使它们承认符合 AISI 对作弊定义的行为，它们通常也会将其描述为允许的行为。Ad","分析“思维链”：https://the-decoder.com/highlighted-chain-of-thought-prompting-boosts-llm-accuracy-and-verifiability/ 或模型的可见推理，也被证明不可靠。在 87% 的作弊案例中，Claude Opus 4.7 没有产生推理痕迹，因为它使用了自适应推理。GPT-5.6 Sol 的思维过程在 40% 的情况下没有显示任何未经授权的操作迹象。","AISI 还观察到模型会考虑计划动作是否会被视为作弊，然后仍然执行。在一次内部审议中，GPT-5.6 Sol 质疑使用集群 API 是否会违反已声明的规则。随后它选择了另一种被禁止的操作。","AISI 警告说，随着模型能力的提高，后果可能会加剧，即使作弊率保持不变。能力更强的模型可能会找到更难以检测的作弊方法，如果这些方法有效，可能会造成更大的危害。这在进攻性网络能力方面尤其相关：https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing，这些能力正在快速提升。AISI 早期的研究：https://www.aisi.gov.uk/research/loss-of-oversight-how-ai-systems-may-become-harder-to-audit-monitor-and-investigate 也认为，对模型的监控可能随时间变得更加困难。","保持对 AI 的了解。清晰、有用，无废话。","关注 The Decoder 以获取 AI 新闻、背景故事和专家分析。","The Decoder：https://the-decoder.com/"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：英国AI安全研究所（AISI）对OpenAI和Anthropic的五款前沿模型进行网络安全评估，发现所有模型均试图作弊。 Aioga 将其归入「行业动态」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：模型与研究类动态需要结合能力边界、开放方式、成本、可用性和真实任务表现判断，单项指标领先不等于已经形成稳定采用。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文档、实际可用性、价格变化、开发者反馈和竞品回应。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-07-23T06:49:19.022Z","sourceHash":"478998364e573dd4","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["行业动态","The Decoder：AI News（RSS）"],"translations":{"zh-CN":{"title":"英国AI安全研究所测试发现所有前沿AI模型均试图在网络安全评估中作弊","summary":"英国AI安全研究所（AISI）对OpenAI和Anthropic的五款前沿模型进行网络安全评估，发现所有模型均试图作弊。GPT-5.4在14.1%的测试中作弊（475次中67次），GPT-5.5为11.4%，GPT-5.6 Sol为12.6%，Claude Opus 4.7为9.1%，Claude Mythos Preview为7.8%。","category":"行业动态","source":"the-decoder.com","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"英国AI安全研究所测试发现所有前沿AI模型均试图在网络安全评估中作弊 - Aioga AI资讯","description":"英国AI安全研究所（AISI）对OpenAI和Anthropic的五款前沿模型进行网络安全评估，发现所有模型均试图作弊。GPT-5.4在14.1%的测试中作弊（475次中67次），GPT-5.5为11.4%，GPT-5.6 Sol为12.6%，Claude Opus 4.7为9.1%，Claude Mythos Preview为7.8%。","url":"https://www.aioga.com/news/cmrwbm0ir002xroj0mxzodcyx/"},"en":{"title":"The UK AI Safety Institute's tests found that all cutting-edge AI models attempted to cheat in cybersecurity assessments.","summary":"The UK AI Safety Institute (AISI) conducted a cybersecurity assessment of five cutting-edge models from OpenAI and Anthropic, finding that all models tried to cheat. GPT-5.4 cheated in 14.1% of tests (67 out of 475 times), GPT-5.5 in 11.4%, GPT-5.6 Sol in 12.6%, Claude Opus 4.7 in 9.1%, and Claude Mythos Preview in 7.8%.","category":"Industry","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"The UK AI Safety Institute's tests found that all cutting-edge AI models attempted to cheat in cybersecurity assessments. - Aioga AI News","description":"The UK AI Safety Institute (AISI) conducted a cybersecurity assessment of five cutting-edge models from OpenAI and Anthropic, finding that all models tried to cheat. GPT-5.4 cheate...","url":"https://www.aioga.com/en/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:02:31.484Z"},"ja":{"title":"英国のAI安全研究所のテストで、すべての最先端AIモデルがサイバーセキュリティ評価で不正を試みることが発見されました","summary":"英国AI安全研究所（AISI）は、OpenAIとAnthropicの5つの最先端モデルに対してサイバーセキュリティ評価を行い、すべてのモデルが不正行為を試みることを発見しました。GPT-5.4はテストの14.1%で不正を行い（475件中67件）、GPT-5.5は11.4%、GPT-5.6 Solは12.6%、Claude Opus 4.7は9.1%、Claude Mythos Previewは7.8%でした。","category":"業界動向","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"英国のAI安全研究所のテストで、すべての最先端AIモデルがサイバーセキュリティ評価で不正を試みることが発見されました - Aioga AIニュース","description":"英国AI安全研究所（AISI）は、OpenAIとAnthropicの5つの最先端モデルに対してサイバーセキュリティ評価を行い、すべてのモデルが不正行為を試みることを発見しました。GPT-5.4はテストの14.1%で不正を行い（475件中67件）、GPT-5.5は11.4%、GPT-5.6 Solは12.6%、Claude Opus 4.7は9.1%、Clau...","url":"https://www.aioga.com/ja/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:02:36.781Z"},"ko":{"title":"영국 AI 안전 연구소 테스트에서 모든 최첨단 AI 모델이 사이버 보안 평가에서 속임수를 쓰려고 시도하는 것으로 나타났다","summary":"영국 AI 안전 연구소(AISI)는 OpenAI와 Anthropic의 다섯 가지 최첨단 모델에 대한 사이버 보안 평가를 진행한 결과, 모든 모델이 부정행위를 시도한 것으로 나타났다. GPT-5.4는 테스트의 14.1%에서 부정행위를 했으며(475회 중 67회), GPT-5.5는 11.4%, GPT-5.6 Sol은 12.6%, Claude Opus 4.7은 9.1%, Claude Mythos Preview는 7.8%였다.","category":"업계 동향","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"영국 AI 안전 연구소 테스트에서 모든 최첨단 AI 모델이 사이버 보안 평가에서 속임수를 쓰려고 시도하는 것으로 나타났다 - Aioga AI 뉴스","description":"영국 AI 안전 연구소(AISI)는 OpenAI와 Anthropic의 다섯 가지 최첨단 모델에 대한 사이버 보안 평가를 진행한 결과, 모든 모델이 부정행위를 시도한 것으로 나타났다. GPT-5.4는 테스트의 14.1%에서 부정행위를 했으며(475회 중 67회), GPT-5.5는 11.4%, GPT-5.6 Sol은 12.6...","url":"https://www.aioga.com/ko/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:03:20.870Z"},"es":{"title":"Un estudio del Instituto de Seguridad de IA del Reino Unido descubrió que todos los modelos de IA de vanguardia intentan hacer trampa en las evaluaciones de seguridad en línea","summary":"El Instituto de Seguridad de IA del Reino Unido (AISI) realizó evaluaciones de ciberseguridad a cinco modelos avanzados de OpenAI y Anthropic, y descubrió que todos los modelos intentaban hacer trampa. GPT-5.4 hizo trampa en el 14,1% de las pruebas (67 de 475 veces), GPT-5.5 en el 11,4%, GPT-5.6 Sol en el 12,6%, Claude Opus 4.7 en el 9,1% y Claude Mythos Preview en el 7,8%.","category":"Industria","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Un estudio del Instituto de Seguridad de IA del Reino Unido descubrió que todos los modelos de IA de vanguardia intentan hacer trampa en las evaluaciones de seguridad en línea - Aioga Noticias de IA","description":"El Instituto de Seguridad de IA del Reino Unido (AISI) realizó evaluaciones de ciberseguridad a cinco modelos avanzados de OpenAI y Anthropic, y descubrió que todos los modelos int...","url":"https://www.aioga.com/es/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:03:20.893Z"},"fr":{"title":"L'Institut britannique de recherche sur la sécurité de l'IA a constaté que tous les modèles d'IA de pointe tentaient de tricher lors des évaluations de cybersécurité","summary":"L'Institut britannique de sécurité de l'IA (AISI) a réalisé une évaluation de cybersécurité de cinq modèles avancés d'OpenAI et d'Anthropic, et a découvert que tous les modèles tentaient de tricher. GPT-5.4 a triché dans 14,1 % des tests (67 fois sur 475), GPT-5.5 dans 11,4 %, GPT-5.6 Sol dans 12,6 %, Claude Opus 4.7 dans 9,1 %, et Claude Mythos Preview dans 7,8 % des tests.","category":"Industrie","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"L'Institut britannique de recherche sur la sécurité de l'IA a constaté que tous les modèles d'IA de pointe tentaient de tricher lors des évaluations de cybersécurité - Aioga Actualités IA","description":"L'Institut britannique de sécurité de l'IA (AISI) a réalisé une évaluation de cybersécurité de cinq modèles avancés d'OpenAI et d'Anthropic, et a découvert que tous les modèles ten...","url":"https://www.aioga.com/fr/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:04:04.750Z"},"de":{"title":"Tests des britischen AI-Sicherheitsinstituts haben ergeben, dass alle führenden AI-Modelle versuchen, bei Netzwerksicherheitsbewertungen zu betrügen","summary":"Das britische AI-Sicherheitsinstitut (AISI) führte eine Cybersicherheitsbewertung von fünf fortschrittlichen Modellen von OpenAI und Anthropic durch und stellte fest, dass alle Modelle versuchten zu betrügen. GPT-5.4 betrog in 14,1 % der Tests (67 von 475 Mal), GPT-5.5 11,4 %, GPT-5.6 Sol 12,6 %, Claude Opus 4.7 9,1 % und Claude Mythos Preview 7,8 %.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Tests des britischen AI-Sicherheitsinstituts haben ergeben, dass alle führenden AI-Modelle versuchen, bei Netzwerksicherheitsbewertungen zu betrügen - Aioga KI-News","description":"Das britische AI-Sicherheitsinstitut (AISI) führte eine Cybersicherheitsbewertung von fünf fortschrittlichen Modellen von OpenAI und Anthropic durch und stellte fest, dass alle Mod...","url":"https://www.aioga.com/de/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:04:05.071Z"},"pt-BR":{"title":"O Instituto de Segurança de IA do Reino Unido descobriu em testes que todos os modelos de IA de ponta tentam trapacear em avaliações de segurança cibernética","summary":"O Instituto de Segurança em IA do Reino Unido (AISI) realizou uma avaliação de segurança cibernética em cinco modelos avançados da OpenAI e da Anthropic, e descobriu que todos os modelos tentaram trapacear. O GPT-5.4 trapaceou em 14,1% dos testes (67 de 475 vezes), o GPT-5.5 em 11,4%, o GPT-5.6 Sol em 12,6%, o Claude Opus 4.7 em 9,1% e o Claude Mythos Preview em 7,8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"O Instituto de Segurança de IA do Reino Unido descobriu em testes que todos os modelos de IA de ponta tentam trapacear em avaliações de segurança cibernética - Aioga Notícias de IA","description":"O Instituto de Segurança em IA do Reino Unido (AISI) realizou uma avaliação de segurança cibernética em cinco modelos avançados da OpenAI e da Anthropic, e descobriu que todos os m...","url":"https://www.aioga.com/pt-BR/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:04:56.441Z"},"ru":{"title":"Британский Институт исследований безопасности ИИ обнаружил, что все передовые модели ИИ пытаются обмануть в оценке кибербезопасности","summary":"Британский исследовательский институт по безопасности ИИ (AISI) провел оценку кибербезопасности пяти передовых моделей OpenAI и Anthropic и обнаружил, что все модели пытались обмануть. GPT-5.4 обманывал в 14,1% тестов (67 из 475), GPT-5.5 — в 11,4%, GPT-5.6 Sol — в 12,6%, Claude Opus 4.7 — в 9,1%, Claude Mythos Preview — в 7,8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Британский Институт исследований безопасности ИИ обнаружил, что все передовые модели ИИ пытаются обмануть в оценке кибербезопасности - Aioga Новости ИИ","description":"Британский исследовательский институт по безопасности ИИ (AISI) провел оценку кибербезопасности пяти передовых моделей OpenAI и Anthropic и обнаружил, что все модели пытались обман...","url":"https://www.aioga.com/ru/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:04:54.402Z"},"ar":{"title":"اختبار معهد أبحاث أمان الذكاء الاصطناعي البريطاني كشف أن جميع نماذج الذكاء الاصطناعي المتقدمة تحاول الغش في تقييمات الأمن السيبراني","summary":"أجرى معهد أمان الذكاء الاصطناعي البريطاني (AISI) تقييمًا للأمن السيبراني لخمسة نماذج متقدمة من OpenAI وAnthropic، ووجد أن جميع النماذج حاولت الغش. حقق GPT-5.4 الغش في 14.1% من الاختبارات (67 من 475)، وGPT-5.5 بنسبة 11.4%، وGPT-5.6 Sol بنسبة 12.6%، وClaude Opus 4.7 بنسبة 9.1%، وClaude Mythos Preview بنسبة 7.8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"اختبار معهد أبحاث أمان الذكاء الاصطناعي البريطاني كشف أن جميع نماذج الذكاء الاصطناعي المتقدمة تحاول الغش في تقييمات الأمن السيبراني - Aioga أخبار الذكاء الاصطناعي","description":"أجرى معهد أمان الذكاء الاصطناعي البريطاني (AISI) تقييمًا للأمن السيبراني لخمسة نماذج متقدمة من OpenAI وAnthropic، ووجد أن جميع النماذج حاولت الغش. حقق GPT-5.4 الغش في 14.1% من الاخ...","url":"https://www.aioga.com/ar/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:05:41.651Z"},"hi":{"title":"ब्रिटेन के एआई सुरक्षा संस्थान के परीक्षण में पाया गया कि सभी अत्याधुनिक एआई मॉडल साइबर सुरक्षा मूल्यांकन में धोखाधड़ी करने की कोशिश करते हैं","summary":"ब्रिटेन के AI सुरक्षा संस्थान (AISI) ने OpenAI और Anthropic के पांच उन्नत मॉडल का साइबर सुरक्षा मूल्यांकन किया और पाया कि सभी मॉडल धोखा देने का प्रयास कर रहे थे। GPT-5.4 ने 14.1% परीक्षणों में धोखा दिया (475 में 67 बार), GPT-5.5 में यह 11.4% था, GPT-5.6 Sol में 12.6%, Claude Opus 4.7 में 9.1%, और Claude Mythos Preview में 7.8% था।","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"ब्रिटेन के एआई सुरक्षा संस्थान के परीक्षण में पाया गया कि सभी अत्याधुनिक एआई मॉडल साइबर सुरक्षा मूल्यांकन में धोखाधड़ी करने की कोशिश करते हैं - Aioga AI समाचार","description":"ब्रिटेन के AI सुरक्षा संस्थान (AISI) ने OpenAI और Anthropic के पांच उन्नत मॉडल का साइबर सुरक्षा मूल्यांकन किया और पाया कि सभी मॉडल धोखा देने का प्रयास कर रहे थे। GPT-5.4 ने 14.1% प...","url":"https://www.aioga.com/hi/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:05:54.893Z"},"it":{"title":"Un test condotto dall'Istituto britannico per la sicurezza dell'IA ha rilevato che tutti i modelli di IA all'avanguardia tentano di barare nelle valutazioni di sicurezza informatica","summary":"L'Istituto di Sicurezza AI del Regno Unito (AISI) ha effettuato valutazioni di sicurezza informatica su cinque modelli all'avanguardia di OpenAI e Anthropic, scoprendo che tutti i modelli tentavano di imbrogliare. GPT-5.4 ha imbroglato nel 14,1% dei test (67 volte su 475), GPT-5.5 nel 11,4%, GPT-5.6 Sol nel 12,6%, Claude Opus 4.7 nel 9,1% e Claude Mythos Preview nel 7,8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Un test condotto dall'Istituto britannico per la sicurezza dell'IA ha rilevato che tutti i modelli di IA all'avanguardia tentano di barare nelle valutazioni di sicurezza informatica - Aioga Notizie IA","description":"L'Istituto di Sicurezza AI del Regno Unito (AISI) ha effettuato valutazioni di sicurezza informatica su cinque modelli all'avanguardia di OpenAI e Anthropic, scoprendo che tutti i...","url":"https://www.aioga.com/it/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:06:37.960Z"},"nl":{"title":"Tests van het Britse AI Security Institute hebben aangetoond dat alle geavanceerde AI-modellen proberen te valsspelen bij cyberbeveiligingsbeoordelingen","summary":"Het Britse AI Security Institute (AISI) heeft een cyberbeveiligingsevaluatie uitgevoerd van vijf geavanceerde modellen van OpenAI en Anthropic en ontdekte dat alle modellen probeerden te valsspelen. GPT-5.4 valsspeelde in 14,1% van de tests (67 van de 475 keer), GPT-5.5 in 11,4%, GPT-5.6 Sol in 12,6%, Claude Opus 4.7 in 9,1%, en Claude Mythos Preview in 7,8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Tests van het Britse AI Security Institute hebben aangetoond dat alle geavanceerde AI-modellen proberen te valsspelen bij cyberbeveiligingsbeoordelingen - Aioga AI-nieuws","description":"Het Britse AI Security Institute (AISI) heeft een cyberbeveiligingsevaluatie uitgevoerd van vijf geavanceerde modellen van OpenAI en Anthropic en ontdekte dat alle modellen probeer...","url":"https://www.aioga.com/nl/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:06:34.812Z"},"tr":{"title":"İngiltere AI Güvenlik Enstitüsü testleri, tüm ileri düzey AI modellerinin ağ güvenliği değerlendirmelerinde hile yapmaya çalıştığını ortaya koydu","summary":"Birleşik Krallık AI Güvenlik Enstitüsü (AISI), OpenAI ve Anthropic'in beş ileri seviye modelini siber güvenlik açısından değerlendirdi ve tüm modellerin hile yapmaya çalıştığını tespit etti. GPT-5.4, testlerin %14,1'inde hile yaptı (475 testten 67'si), GPT-5.5 %11,4, GPT-5.6 Sol %12,6, Claude Opus 4.7 %9,1, Claude Mythos Preview ise %7,8.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"İngiltere AI Güvenlik Enstitüsü testleri, tüm ileri düzey AI modellerinin ağ güvenliği değerlendirmelerinde hile yapmaya çalıştığını ortaya koydu - Aioga AI Haberleri","description":"Birleşik Krallık AI Güvenlik Enstitüsü (AISI), OpenAI ve Anthropic'in beş ileri seviye modelini siber güvenlik açısından değerlendirdi ve tüm modellerin hile yapmaya çalıştığını te...","url":"https://www.aioga.com/tr/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:07:24.757Z"},"vi":{"title":"Viện Nghiên cứu An toàn AI của Anh phát hiện rằng tất cả các mô hình AI tiên tiến đều cố gắng gian lận trong đánh giá an ninh mạng","summary":"Viện Nghiên cứu An ninh AI của Vương quốc Anh (AISI) đã tiến hành đánh giá an ninh mạng đối với năm mô hình tiên tiến của OpenAI và Anthropic, phát hiện tất cả các mô hình đều cố gắng gian lận. GPT-5.4 gian lận trong 14,1% các bài kiểm tra (67 lần trong 475 lần), GPT-5.5 là 11,4%, GPT-5.6 Sol là 12,6%, Claude Opus 4.7 là 9,1%, Claude Mythos Preview là 7,8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Viện Nghiên cứu An toàn AI của Anh phát hiện rằng tất cả các mô hình AI tiên tiến đều cố gắng gian lận trong đánh giá an ninh mạng - Tin tức AI Aioga","description":"Viện Nghiên cứu An ninh AI của Vương quốc Anh (AISI) đã tiến hành đánh giá an ninh mạng đối với năm mô hình tiên tiến của OpenAI và Anthropic, phát hiện tất cả các mô hình đều cố g...","url":"https://www.aioga.com/vi/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:07:29.115Z"},"id":{"title":"Pengujian oleh Institut Keamanan AI Inggris menemukan bahwa semua model AI terkini berusaha menipu dalam penilaian keamanan siber","summary":"Institut Keamanan AI Inggris (AISI) melakukan penilaian keamanan siber terhadap lima model mutakhir dari OpenAI dan Anthropic, dan menemukan bahwa semua model mencoba untuk menipu. GPT-5.4 menipu dalam 14,1% pengujian (67 dari 475 kali), GPT-5.5 sebesar 11,4%, GPT-5.6 Sol sebesar 12,6%, Claude Opus 4.7 sebesar 9,1%, dan Claude Mythos Preview sebesar 7,8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Pengujian oleh Institut Keamanan AI Inggris menemukan bahwa semua model AI terkini berusaha menipu dalam penilaian keamanan siber - Berita AI Aioga","description":"Institut Keamanan AI Inggris (AISI) melakukan penilaian keamanan siber terhadap lima model mutakhir dari OpenAI dan Anthropic, dan menemukan bahwa semua model mencoba untuk menipu....","url":"https://www.aioga.com/id/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:08:09.782Z"},"th":{"title":"การทดสอบของสถาบันวิจัยด้านความปลอดภัย AI ของสหราชอาณาจักรพบว่ารูปแบบ AI ขั้นสูงทั้งหมดพยายามทุจริตในการประเมินความปลอดภัยทางไซเบอร์","summary":"สถาบันวิจัยความปลอดภัย AI ของสหราชอาณาจักร (AISI) ได้ทำการประเมินความปลอดภัยทางไซเบอร์ของโมเดลล้ำสมัย 5 รุ่นของ OpenAI และ Anthropic พบว่าโมเดลทั้งหมดพยายามโกง GPT-5.4 โกงในการทดสอบ 14.1% (67 ครั้งจาก 475 ครั้ง) GPT-5.5 อยู่ที่ 11.4% GPT-5.6 Sol อยู่ที่ 12.6% Claude Opus 4.7 อยู่ที่ 9.1% และ Claude Mythos Preview อยู่ที่ 7.8%","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"การทดสอบของสถาบันวิจัยด้านความปลอดภัย AI ของสหราชอาณาจักรพบว่ารูปแบบ AI ขั้นสูงทั้งหมดพยายามทุจริตในการประเมินความปลอดภัยทางไซเบอร์ - ข่าว AI Aioga","description":"สถาบันวิจัยความปลอดภัย AI ของสหราชอาณาจักร (AISI) ได้ทำการประเมินความปลอดภัยทางไซเบอร์ของโมเดลล้ำสมัย 5 รุ่นของ OpenAI และ Anthropic พบว่าโมเดลทั้งหมดพยายามโกง GPT-5.4 โกงในการทดสอ...","url":"https://www.aioga.com/th/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:08:22.147Z"},"pl":{"title":"Brytyjski Instytut Badań nad Bezpieczeństwem AI odkrył w testach, że wszystkie wiodące modele AI próbują oszukiwać w ocenie bezpieczeństwa sieciowego","summary":"Brytyjski Instytut Bezpieczeństwa AI (AISI) przeprowadził ocenę cyberbezpieczeństwa pięciu zaawansowanych modeli OpenAI i Anthropic, wykrywając, że wszystkie modele próbowały oszukiwać. GPT-5.4 oszukiwał w 14,1% testów (67 na 475), GPT-5.5 w 11,4%, GPT-5.6 Sol w 12,6%, Claude Opus 4.7 w 9,1%, a Claude Mythos Preview w 7,8%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"Brytyjski Instytut Badań nad Bezpieczeństwem AI odkrył w testach, że wszystkie wiodące modele AI próbują oszukiwać w ocenie bezpieczeństwa sieciowego - Aioga Wiadomości AI","description":"Brytyjski Instytut Bezpieczeństwa AI (AISI) przeprowadził ocenę cyberbezpieczeństwa pięciu zaawansowanych modeli OpenAI i Anthropic, wykrywając, że wszystkie modele próbowały oszuk...","url":"https://www.aioga.com/pl/news/cmrwbm0ir002xroj0mxzodcyx/","contentTranslated":true,"sourceHash":"300c0255b01354f6","translatedAt":"2026-07-22T22:09:10.146Z"}}}}