{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-28T06:03:00.468Z","headline":"GPT-6 Astra 幻觉更少但仍易受隐藏提示词注入攻击","description":"The Decoder 报道，OpenAI 新模型 GPT-6 Astra 幻觉少于前代 GPT-5.6 Sol，直接提示词注入防御率达 99.99%，但多轮自适应攻击下防御率降至约 67%。","url":"https://www.aioga.com/news/cmtn8fc1w0qb4romyobllbzv9/","mainEntityOfPage":"https://www.aioga.com/news/cmtn8fc1w0qb4romyobllbzv9/","datePublished":"2026-09-04T17:23:35.000Z","dateModified":"2026-09-04T17:23:35.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://the-decoder.com/openais-gpt-6-astra-hallucinates-less-but-remains-vulnerable-to-hidden-prompt-injections","https://aihot.news/items/cmtn8fc1w0qb4romyobllbzv9"],"canonicalUrl":"https://www.aioga.com/news/cmtn8fc1w0qb4romyobllbzv9/","directAnswer":{"@type":"Answer","text":"The Decoder 报道称，OpenAI 的 GPT-6 Astra 在被用户标记为错误回答的 ChatGPT 对话测试中，比 GPT-5.6 Sol 更少复现事实错误。其对直接提示词注入的防御率为 99.99%，但在多轮自适应攻击测试中约降至 67%。","url":"https://www.aioga.com/news/cmtn8fc1w0qb4romyobllbzv9/","dateCreated":"2026-09-04T17:23:35.000Z","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/openais-gpt-6-astra-hallucinates-less-but-remains-vulnerable-to-hidden-prompt-injections","datePublished":"2026-09-04T17:23:35.000Z","provider":{"@type":"Organization","name":"the-decoder.com","url":"https://the-decoder.com/openais-gpt-6-astra-hallucinates-less-but-remains-vulnerable-to-hidden-prompt-injections"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/cmtn8fc1w0qb4romyobllbzv9","datePublished":"2026-09-04T17:23:35.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/cmtn8fc1w0qb4romyobllbzv9"}}],"aggregationSource":"The Decoder：AI News（RSS）","originalPublisher":{"name":"the-decoder.com","url":"https://the-decoder.com/openais-gpt-6-astra-hallucinates-less-but-remains-vulnerable-to-hidden-prompt-injections"},"geoDeepAnswer":null,"article":{"id":"cmtn8fc1w0qb4romyobllbzv9","slug":"cmtn8fc1w0qb4romyobllbzv9","url":"https://www.aioga.com/news/cmtn8fc1w0qb4romyobllbzv9/","title":"GPT-6 Astra 幻觉更少但仍易受隐藏提示词注入攻击","title_en":"","summary":"The Decoder 报道，OpenAI 新模型 GPT-6 Astra 幻觉少于前代 GPT-5.6 Sol，直接提示词注入防御率达 99.99%，但多轮自适应攻击下防御率降至约 67%。","source":"The Decoder：AI News（RSS）","sourceUrl":"https://the-decoder.com/openais-gpt-6-astra-hallucinates-less-but-remains-vulnerable-to-hidden-prompt-injections","aiHotUrl":"https://aihot.news/items/cmtn8fc1w0qb4romyobllbzv9","publishedAt":"2026-09-04T17:23:35.000Z","category":"行业动态","score":72,"selected":true,"articleBody":["OpenAI's new model, GPT-6 Astra, produces fewer hallucinations and blocks prompt injection attacks more effectively than its predecessors. But it still isn't reliable enough for truly secure AI agent deployments.","The new Astra model：https://the-decoder.com/benchmarks-disagree-on-gpt-6-astra-but-its-human-beating-efficiency-on-arc-agi-3-pulls-chollets-agi-forecast-forward/ makes far fewer factual errors than its predecessor, GPT-5.6 Sol, according to OpenAI's system card：https://deploymentsafety.openai.com/gpt-6-astra/. OpenAI tested it against ChatGPT conversations that users had flagged for wrong answers, meaning these were particularly error-prone cases whose failure rates shouldn't be taken as typical for everyday use. Astra reproduced these reported errors much less often, with the biggest gains showing up at low latency settings and lower reasoning levels.","For direct prompt injections, where users try to manipulate the model through their own prompts, Astra hits a near-perfect 99.99 percent defense rate. OpenAI credits its GPT-Red method：https://the-decoder.com/openai-is-now-using-ai-to-attack-its-own-ai-and-its-working-better-than-humans-ever-did/ for this, which uses an automated attacker to harden the model during training. Ad","Jailbreak resistance looks similar. Against a fixed dataset of known attacks trying to extract harmful responses about biology, violence, and cybersecurity, Astra refuses to help in 91.5 to 98.3 percent of cases. Ad","When attackers adapt their strategy over multiple conversation rounds, Astra's defense rate drops to about 67 percent, meaning persistent adversaries can coax out at least one problematic response roughly one in three tries. Predecessor models scored just under 50 percent on the same test. OpenAI notes that these tests ran on the bare model without the production safety layers like classifiers that ship with the actual product.","Astra makes progress on indirect prompt injections, where an attack is buried inside a document the AI reads. External testing by security firm Gray Swan, using 1,810 curated attacks from their IPI Arena：https://arxiv.org/abs/2603.15714, found that with 15 attempts per scenario, Astra was cracked at least once 8.5 percent of the time. GPT-5.6 Sol failed 27 percent of the time. Claude Opus 5 did better at 4.8 percent in the same evaluation, but it wasn't immune either. Ad","The numbers in Gray Swan's combined Q1 and Q2 test actually went up compared to earlier results. Anthropic previously reported only a two percent attack success rate：https://the-decoder.com/opus-5-may-have-solved-browser-based-prompt-injection-the-biggest-security-flaw-haunting-ai-agents/ based on the easier Q1 test alone, and GPT-5.6 Sol scored just 20 percent there too. Anthropic also ran all models with extended reasoning turned on, which, alongside the broader test scope, could explain the gap.","Even though these are curated, hand-picked attacks, the success rates should worry any enterprise security team. Astra can be tricked through injected instructions in roughly one out of every twelve scenarios. Opus 5 holds up better, but it still fails about one in twenty-one. And the risk is growing. AI agents are increasingly writing code, operating tools, and controlling computers on their own, which is what Gray Swan tested. These agents are also being built to run around the clock and at scale, with reading and processing documents as one of their core jobs. Ad","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/2026/09/gpt6_hallucination_rate.png","alt":"","afterParagraph":1,"url":"/media/articles/cmtn8fc1w0qb4romyobllbzv9/060d31c08e0ac9d8.png"},{"sourceUrl":"https://the-decoder.com/wp-content/uploads/2026/09/prompt_injection_grayswan.png","alt":"","afterParagraph":5,"url":"/media/articles/cmtn8fc1w0qb4romyobllbzv9/6d6e92227464e9e6.png"}],"mediaStatus":"ok","articleBodyZh":["OpenAI的新模型GPT-6 Astra比其前代产品产生的虚假信息更少，并且能够更有效地阻止提示注入攻击。但它仍然不足以用于真正安全的AI代理部署。","根据OpenAI的系统卡:https://deploymentsafety.openai.com/gpt-6-astra/，新的Astra模型：https://the-decoder.com/benchmarks-disagree-on-gpt-6-astra-but-its-human-beating-efficiency-on-arc-agi-3-pulls-chollets-agi-forecast-forward/比其前代模型GPT-5.6 Sol犯的事实性错误要少得多。OpenAI将其测试对象设定为用户标记出错误答案的ChatGPT对话，这意味着这些案例特别容易出错，其失败率不应被视为日常使用的典型情况。Astra重现这些报告错误的频率明显下降，最大改进出现在低延迟设置和较低推理水平下。","对于直接提示注入，即用户试图通过自己的提示操控模型，Astra的防御率接近完美，达到99.99%。OpenAI将此归功于其GPT-Red方法：https://the-decoder.com/openai-is-now-using-ai-to-attack-its-own-ai-and-its-working-better-than-humans-ever-did/，该方法使用自动化攻击者在训练过程中强化模型。Ad","防越狱能力表现类似。在尝试提取有关生物学、暴力和网络安全的有害响应的已知攻击固定数据集上，Astra在91.5%到98.3%的案例中拒绝提供帮助。Ad","当攻击者在多轮对话中调整策略时，Astra的防御率下降到约67%，这意味着持续的对手大约每三次尝试中就能诱导出一次有问题的响应。同样的测试中，前代模型的得分刚低于50%。OpenAI指出，这些测试是在裸模型上进行的，没有使用实际产品中附带的分类器等生产安全层。","Astra在间接提示注入方面取得进展，这种攻击被隐藏在AI读取的文档中。安全公司Gray Swan的外部测试中，使用他们的IPI Arena收集的1,810个攻击案例：https://arxiv.org/abs/2603.15714，发现在每种场景尝试15次时，Astra至少有一次被攻破的概率为8.5%。GPT-5.6 Sol的失败率为27%。Claude Opus 5在同一评估中表现更好，失败率为4.8%，但也并非完全免疫。广告","Gray Swan结合Q1和Q2的测试数据显示，数字实际上比早期结果还要高。Anthropic之前仅基于较简单的Q1测试报告了2%的攻击成功率：https://the-decoder.com/opus-5-may-have-solved-browser-based-prompt-injection-the-biggest-security-flaw-haunting-ai-agents/，而GPT-5.6 Sol在那里的得分也只有20%。Anthropic还在所有模型上启用了扩展推理功能，这与更广泛的测试范围一起，可能解释了差距。","即使这些攻击是经过策划和人工挑选的，其成功率也应让任何企业安全团队感到担忧。Astra在大约每十二个场景中就可能被注入指令欺骗一次。Opus 5表现更好，但仍大约有每二十一次失败一次的情况。而且风险还在增长。AI代理正越来越多地编写代码、操作工具并独立控制计算机，这也是Gray Swan所测试的内容。这些代理还被设计为全天候、大规模运行，将读取和处理文档作为核心工作之一。广告","保持对人工智能的了解。清晰、有用，无废话。","关注The Decoder获取AI新闻、背景故事和专家分析。","解码器：https://the-decoder.com/"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"The Decoder 报道称，OpenAI 的 GPT-6 Astra 在被用户标记为错误回答的 ChatGPT 对话测试中，比 GPT-5.6 Sol 更少复现事实错误。其对直接提示词注入的防御率为 99.99%，但在多轮自适应攻击测试中约降至 67%。","background":"来源称，这些幻觉测试选取的是用户曾标记为错误的对话，因此失败率不应视为日常使用的典型水平。安全公司 Gray Swan 以 1,810 个精选间接提示词注入攻击测试，在每个场景尝试 15 次时，Astra 至少一次被攻破的比例为 8.5%。","viewpoint":"Aioga 判断：Astra 在来源披露的错误复现、直接注入和间接注入测试中均显示出相较 GPT-5.6 Sol 的改进，但多轮自适应攻击与间接注入结果表明，其安全表现仍不足以单独证明智能体部署已具备可靠的安全性。","implications":"可能影响：计划让模型读取文档、调用工具或操作计算机的使用方，可能需要继续保留产品级安全层与多轮攻击评估。来源中的测试为特定错误对话和精选攻击集，结果不代表所有真实场景，也不足以推导实际部署中的总体风险。","nextStep":"后续观察：可关注 OpenAI 是否披露更多覆盖真实部署条件的测试，以及其生产环境安全层在多轮自适应和间接提示词注入情境下的表现。也值得关注不同测试范围、尝试次数与推理设置对结果可比性的影响。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-09-04T18:26:53.421Z","sourceHash":"843dcaa27b7f1f58","review":{"approved":true,"groundedness":94,"clarity":92,"duplicationRisk":20,"blockingIssues":[],"notes":["候选内容准确保留了来源中的关键限定：幻觉测试来自用户标记错误的对话，不代表日常典型失败率。","Gray Swan 测试的攻击数量、15 次尝试、8.5% 被攻破比例均与来源一致。","viewpoint 和 implications 使用了“判断”“可能”“不足以推导”等限定语，基本没有把推论冒充为确定事实。","可选优化：viewpoint 中“直接注入测试中相较 GPT-5.6 Sol 的改进”来源对直接注入更多是概括性称优于前代，若追求更严谨可改为“相较前代模型”。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","The Decoder：AI News（RSS）"],"translations":{"zh-CN":{"title":"GPT-6 Astra 幻觉更少但仍易受隐藏提示词注入攻击","summary":"The Decoder 报道，OpenAI 新模型 GPT-6 Astra 幻觉少于前代 GPT-5.6 Sol，直接提示词注入防御率达 99.99%，但多轮自适应攻击下防御率降至约 67%。","category":"行业动态","source":"the-decoder.com","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra 幻觉更少但仍易受隐藏提示词注入攻击 - Aioga AI资讯","description":"The Decoder 报道，OpenAI 新模型 GPT-6 Astra 幻觉少于前代 GPT-5.6 Sol，直接提示词注入防御率达 99.99%，但多轮自适应攻击下防御率降至约 67%。","url":"https://www.aioga.com/news/cmtn8fc1w0qb4romyobllbzv9/","articleBody":["OpenAI的新模型GPT-6 Astra比其前代产品产生的虚假信息更少，并且能够更有效地阻止提示注入攻击。但它仍然不足以用于真正安全的AI代理部署。","根据OpenAI的系统卡:https://deploymentsafety.openai.com/gpt-6-astra/，新的Astra模型：https://the-decoder.com/benchmarks-disagree-on-gpt-6-astra-but-its-human-beating-efficiency-on-arc-agi-3-pulls-chollets-agi-forecast-forward/比其前代模型GPT-5.6 Sol犯的事实性错误要少得多。OpenAI将其测试对象设定为用户标记出错误答案的ChatGPT对话，这意味着这些案例特别容易出错，其失败率不应被视为日常使用的典型情况。Astra重现这些报告错误的频率明显下降，最大改进出现在低延迟设置和较低推理水平下。","对于直接提示注入，即用户试图通过自己的提示操控模型，Astra的防御率接近完美，达到99.99%。OpenAI将此归功于其GPT-Red方法：https://the-decoder.com/openai-is-now-using-ai-to-attack-its-own-ai-and-its-working-better-than-humans-ever-did/，该方法使用自动化攻击者在训练过程中强化模型。Ad","防越狱能力表现类似。在尝试提取有关生物学、暴力和网络安全的有害响应的已知攻击固定数据集上，Astra在91.5%到98.3%的案例中拒绝提供帮助。Ad","当攻击者在多轮对话中调整策略时，Astra的防御率下降到约67%，这意味着持续的对手大约每三次尝试中就能诱导出一次有问题的响应。同样的测试中，前代模型的得分刚低于50%。OpenAI指出，这些测试是在裸模型上进行的，没有使用实际产品中附带的分类器等生产安全层。","Astra在间接提示注入方面取得进展，这种攻击被隐藏在AI读取的文档中。安全公司Gray Swan的外部测试中，使用他们的IPI Arena收集的1,810个攻击案例：https://arxiv.org/abs/2603.15714，发现在每种场景尝试15次时，Astra至少有一次被攻破的概率为8.5%。GPT-5.6 Sol的失败率为27%。Claude Opus 5在同一评估中表现更好，失败率为4.8%，但也并非完全免疫。广告","Gray Swan结合Q1和Q2的测试数据显示，数字实际上比早期结果还要高。Anthropic之前仅基于较简单的Q1测试报告了2%的攻击成功率：https://the-decoder.com/opus-5-may-have-solved-browser-based-prompt-injection-the-biggest-security-flaw-haunting-ai-agents/，而GPT-5.6 Sol在那里的得分也只有20%。Anthropic还在所有模型上启用了扩展推理功能，这与更广泛的测试范围一起，可能解释了差距。","即使这些攻击是经过策划和人工挑选的，其成功率也应让任何企业安全团队感到担忧。Astra在大约每十二个场景中就可能被注入指令欺骗一次。Opus 5表现更好，但仍大约有每二十一次失败一次的情况。而且风险还在增长。AI代理正越来越多地编写代码、操作工具并独立控制计算机，这也是Gray Swan所测试的内容。这些代理还被设计为全天候、大规模运行，将读取和处理文档作为核心工作之一。广告","保持对人工智能的了解。清晰、有用，无废话。","关注The Decoder获取AI新闻、背景故事和专家分析。","解码器：https://the-decoder.com/"]},"en":{"title":"GPT-6 Astra has fewer hallucinations but is still susceptible to hidden prompt injection attacks","summary":"The Decoder reported that OpenAI's new model GPT-6 Astra has fewer hallucinations than its predecessor GPT-5.6 Sol, with a direct prompt injection defense rate of 99.99%, but under multi-turn adaptive attacks, the defense rate drops to about 67%.","category":"Industry","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra has fewer hallucinations but is still susceptible to hidden prompt injection attacks - Aioga AI News","description":"The Decoder reported that OpenAI's new model GPT-6 Astra has fewer hallucinations than its predecessor GPT-5.6 Sol, with a direct prompt injection defense rate of 99.99%, but under...","url":"https://www.aioga.com/en/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:42:29.375Z"},"ja":{"title":"GPT-6 Astra 幻覚は少ないが、依然として隠れたプロンプト注入攻撃を受けやすい","summary":"The Decoder の報道によると、OpenAI の新モデル GPT-6 Astra は前世代の GPT-5.6 Sol よりも幻覚が少なく、直接プロンプトインジェクションの防御率は 99.99% に達するものの、多回自適応攻撃の下では防御率が約 67% に低下する。","category":"業界動向","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra 幻覚は少ないが、依然として隠れたプロンプト注入攻撃を受けやすい - Aioga AIニュース","description":"The Decoder の報道によると、OpenAI の新モデル GPT-6 Astra は前世代の GPT-5.6 Sol よりも幻覚が少なく、直接プロンプトインジェクションの防御率は 99.99% に達するものの、多回自適応攻撃の下では防御率が約 67% に低下する。","url":"https://www.aioga.com/ja/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:42:38.098Z"},"ko":{"title":"GPT-6 Astra 환각이 더 적지만 여전히 숨겨진 프롬프트 주입 공격에 취약함","summary":"The Decoder 보도에 따르면, OpenAI의 새로운 모델 GPT-6 Astra는 이전 세대 GPT-5.6 Sol보다 환각 현상이 적고, 직접적인 프롬프트 주입 방어율은 99.99%에 달하지만, 다중 라운드 적응 공격에서는 방어율이 약 67%로 내려간다.","category":"업계 동향","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra 환각이 더 적지만 여전히 숨겨진 프롬프트 주입 공격에 취약함 - Aioga AI 뉴스","description":"The Decoder 보도에 따르면, OpenAI의 새로운 모델 GPT-6 Astra는 이전 세대 GPT-5.6 Sol보다 환각 현상이 적고, 직접적인 프롬프트 주입 방어율은 99.99%에 달하지만, 다중 라운드 적응 공격에서는 방어율이 약 67%로 내려간다.","url":"https://www.aioga.com/ko/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:43:18.102Z"},"es":{"title":"GPT-6 Astra tiene menos alucinaciones pero sigue siendo susceptible a ataques de inyección de prompts ocultos","summary":"The Decoder informó que el nuevo modelo de OpenAI, GPT-6 Astra, tiene menos alucinaciones que su predecesor GPT-5.6 Sol, y la tasa de defensa contra la inyección directa de prompts alcanza el 99,99%, pero bajo ataques adaptativos de múltiples rondas, la tasa de defensa cae aproximadamente al 67%.","category":"Industria","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra tiene menos alucinaciones pero sigue siendo susceptible a ataques de inyección de prompts ocultos - Aioga Noticias de IA","description":"The Decoder informó que el nuevo modelo de OpenAI, GPT-6 Astra, tiene menos alucinaciones que su predecesor GPT-5.6 Sol, y la tasa de defensa contra la inyección directa de prompts...","url":"https://www.aioga.com/es/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:43:15.191Z"},"fr":{"title":"GPT-6 Astra a moins d'hallucinations mais reste vulnérable aux attaques par injection de mots-clés cachés","summary":"The Decoder a rapporté que le nouveau modèle GPT-6 Astra d'OpenAI a moins d'hallucinations que son prédécesseur GPT-5.6 Sol, avec un taux de défense contre l'injection directe de mots-clés atteignant 99,99 %, mais ce taux de défense tombe à environ 67 % lors d'attaques adaptatives en plusieurs tours.","category":"Industrie","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra a moins d'hallucinations mais reste vulnérable aux attaques par injection de mots-clés cachés - Aioga Actualités IA","description":"The Decoder a rapporté que le nouveau modèle GPT-6 Astra d'OpenAI a moins d'hallucinations que son prédécesseur GPT-5.6 Sol, avec un taux de défense contre l'injection directe de m...","url":"https://www.aioga.com/fr/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:43:55.987Z"},"de":{"title":"GPT-6 Astra hat weniger Halluzinationen, ist aber immer noch anfällig für Angriffe durch versteckte Prompt-Injektionen","summary":"The Decoder berichtet, dass das neue OpenAI-Modell GPT-6 Astra weniger Halluzinationen als sein Vorgänger GPT-5.6 Sol aufweist. Die Abwehrrate gegen direkte Prompt-Injektionen beträgt 99,99%, aber bei mehrstufigen adaptiven Angriffen sinkt die Abwehrrate auf etwa 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra hat weniger Halluzinationen, ist aber immer noch anfällig für Angriffe durch versteckte Prompt-Injektionen - Aioga KI-News","description":"The Decoder berichtet, dass das neue OpenAI-Modell GPT-6 Astra weniger Halluzinationen als sein Vorgänger GPT-5.6 Sol aufweist. Die Abwehrrate gegen direkte Prompt-Injektionen betr...","url":"https://www.aioga.com/de/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:43:55.176Z"},"pt-BR":{"title":"GPT-6 Astra tem menos alucinações, mas ainda é suscetível a ataques de injeção de prompts ocultos","summary":"O The Decoder relatou que o novo modelo da OpenAI, GPT-6 Astra, apresenta menos alucinações do que o antecessor GPT-5.6 Sol, com uma taxa de defesa contra injeção direta de prompts de 99,99%, mas sob ataques adaptativos de múltiplas rodadas a taxa de defesa cai para cerca de 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra tem menos alucinações, mas ainda é suscetível a ataques de injeção de prompts ocultos - Aioga Notícias de IA","description":"O The Decoder relatou que o novo modelo da OpenAI, GPT-6 Astra, apresenta menos alucinações do que o antecessor GPT-5.6 Sol, com uma taxa de defesa contra injeção direta de prompts...","url":"https://www.aioga.com/pt-BR/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:44:33.141Z"},"ru":{"title":"GPT-6 Astra вызывает меньше галлюцинаций, но по-прежнему уязвим к атакам с использованием скрытых подсказок","summary":"The Decoder сообщает, что у новой модели OpenAI GPT-6 Astra меньше галлюцинаций, чем у предыдущей GPT-5.6 Sol, и уровень защиты от прямого внедрения подсказок достигает 99,99%, но при многоэтапных адаптивных атаках уровень защиты снижается примерно до 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra вызывает меньше галлюцинаций, но по-прежнему уязвим к атакам с использованием скрытых подсказок - Aioga Новости ИИ","description":"The Decoder сообщает, что у новой модели OpenAI GPT-6 Astra меньше галлюцинаций, чем у предыдущей GPT-5.6 Sol, и уровень защиты от прямого внедрения подсказок достигает 99,99%, но...","url":"https://www.aioga.com/ru/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:44:36.423Z"},"ar":{"title":"جي بي تي-6 أسترا يسبب هلوسة أقل ولكنه لا يزال عرضة لهجمات حقن تعليمات خفية","summary":"ذكرت The Decoder أن نموذج OpenAI الجديد GPT-6 Astra لديه أوهام أقل مقارنة بالجيل السابق GPT-5.6 Sol، وأن معدل الدفاع ضد حقن الإرشادات المباشرة يبلغ 99.99٪، لكن في ظل الهجمات التكيفية متعددة الجولات ينخفض معدل الدفاع إلى حوالي 67٪.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"جي بي تي-6 أسترا يسبب هلوسة أقل ولكنه لا يزال عرضة لهجمات حقن تعليمات خفية - Aioga أخبار الذكاء الاصطناعي","description":"ذكرت The Decoder أن نموذج OpenAI الجديد GPT-6 Astra لديه أوهام أقل مقارنة بالجيل السابق GPT-5.6 Sol، وأن معدل الدفاع ضد حقن الإرشادات المباشرة يبلغ 99.99٪، لكن في ظل الهجمات التكيف...","url":"https://www.aioga.com/ar/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:45:13.979Z"},"hi":{"title":"GPT-6 Astra में भ्रम कम हैं लेकिन यह अभी भी छिपे हुए प्रॉम्प्ट इंजेक्शन हमलों के प्रति संवेदनशील है","summary":"The Decoder ने बताया कि OpenAI का नया मॉडल GPT-6 Astra में भूतपूर्व GPT-5.6 Sol की तुलना में भ्रम कम है, सीधे प्रॉम्प्ट इंजेक्शन डिफेंस दर 99.99% तक है, लेकिन कई दौर की आत्म-अनुकूलित हमलों में यह डिफेंस दर लगभग 67% तक गिर जाती है।","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra में भ्रम कम हैं लेकिन यह अभी भी छिपे हुए प्रॉम्प्ट इंजेक्शन हमलों के प्रति संवेदनशील है - Aioga AI समाचार","description":"The Decoder ने बताया कि OpenAI का नया मॉडल GPT-6 Astra में भूतपूर्व GPT-5.6 Sol की तुलना में भ्रम कम है, सीधे प्रॉम्प्ट इंजेक्शन डिफेंस दर 99.99% तक है, लेकिन कई दौर की आत्म-अनुकूल...","url":"https://www.aioga.com/hi/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:45:16.675Z"},"it":{"title":"GPT-6 Astra ha meno allucinazioni ma è ancora vulnerabile agli attacchi di iniezione con prompt nascosti","summary":"The Decoder riporta che il nuovo modello di OpenAI, GPT-6 Astra, ha meno allucinazioni rispetto alla precedente generazione GPT-5.6 Sol, con un tasso di difesa contro l'iniezione diretta di prompt del 99,99%, ma sotto attacchi adattivi multi-turno il tasso di difesa scende a circa il 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra ha meno allucinazioni ma è ancora vulnerabile agli attacchi di iniezione con prompt nascosti - Aioga Notizie IA","description":"The Decoder riporta che il nuovo modello di OpenAI, GPT-6 Astra, ha meno allucinazioni rispetto alla precedente generazione GPT-5.6 Sol, con un tasso di difesa contro l'iniezione d...","url":"https://www.aioga.com/it/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:45:54.074Z"},"nl":{"title":"GPT-6 Astra heeft minder hallucinaties maar is nog steeds vatbaar voor verborgen promptinjectieaanvallen","summary":"The Decoder meldde dat het nieuwe model van OpenAI, GPT-6 Astra, minder hallucinaties heeft dan de vorige generatie GPT-5.6 Sol, met een verdediging tegen directe promptinjecties van 99,99%, maar onder meervoudige adaptieve aanvallen daalt de verdedigingsgraad tot ongeveer 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra heeft minder hallucinaties maar is nog steeds vatbaar voor verborgen promptinjectieaanvallen - Aioga AI-nieuws","description":"The Decoder meldde dat het nieuwe model van OpenAI, GPT-6 Astra, minder hallucinaties heeft dan de vorige generatie GPT-5.6 Sol, met een verdediging tegen directe promptinjecties v...","url":"https://www.aioga.com/nl/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:45:55.318Z"},"tr":{"title":"GPT-6 Astra daha az halüsinasyon yapıyor ancak hâlâ gizli istem kelimelerine yönelik saldırılara açıktır","summary":"The Decoder, OpenAI'ın yeni modeli GPT-6 Astra'nın selefi GPT-5.6 Sol'a kıyasla daha az halüsinasyon ürettiğini, doğrudan istem enjeksiyonlarına karşı savunma oranının %99,99 olduğunu, ancak çok turlu adaptif saldırılar altında savunma oranının yaklaşık %67'ye düştüğünü bildirdi.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra daha az halüsinasyon yapıyor ancak hâlâ gizli istem kelimelerine yönelik saldırılara açıktır - Aioga AI Haberleri","description":"The Decoder, OpenAI'ın yeni modeli GPT-6 Astra'nın selefi GPT-5.6 Sol'a kıyasla daha az halüsinasyon ürettiğini, doğrudan istem enjeksiyonlarına karşı savunma oranının %99,99 olduğ...","url":"https://www.aioga.com/tr/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:46:34.647Z"},"vi":{"title":"GPT-6 Astra ảo giác ít hơn nhưng vẫn dễ bị tấn công tiêm nhiễm từ khóa ẩn","summary":"The Decoder đưa tin, mô hình mới GPT-6 Astra của OpenAI ít ảo giác hơn so với thế hệ trước GPT-5.6 Sol, tỷ lệ phòng thủ chống chèn trực tiếp từ từ khóa đạt 99,99%, nhưng dưới các cuộc tấn công thích ứng nhiều vòng, tỷ lệ phòng thủ giảm xuống khoảng 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra ảo giác ít hơn nhưng vẫn dễ bị tấn công tiêm nhiễm từ khóa ẩn - Tin tức AI Aioga","description":"The Decoder đưa tin, mô hình mới GPT-6 Astra của OpenAI ít ảo giác hơn so với thế hệ trước GPT-5.6 Sol, tỷ lệ phòng thủ chống chèn trực tiếp từ từ khóa đạt 99,99%, nhưng dưới các c...","url":"https://www.aioga.com/vi/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:46:30.449Z"},"id":{"title":"GPT-6 Astra memiliki halusinasi lebih sedikit tetapi masih rentan terhadap serangan injeksi prompt tersembunyi","summary":"The Decoder melaporkan, model baru OpenAI GPT-6 Astra memiliki halusinasi lebih sedikit dibandingkan pendahulunya GPT-5.6 Sol, tingkat pertahanan terhadap injeksi prompt langsung mencapai 99,99%, tetapi di bawah serangan adaptif multi-putaran tingkat pertahanannya turun menjadi sekitar 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra memiliki halusinasi lebih sedikit tetapi masih rentan terhadap serangan injeksi prompt tersembunyi - Berita AI Aioga","description":"The Decoder melaporkan, model baru OpenAI GPT-6 Astra memiliki halusinasi lebih sedikit dibandingkan pendahulunya GPT-5.6 Sol, tingkat pertahanan terhadap injeksi prompt langsung m...","url":"https://www.aioga.com/id/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:47:10.773Z"},"th":{"title":"GPT-6 Astra มีอาการประสาทหลอนน้อยลงแต่ยังคงเสี่ยงต่อการถูกโจมตีด้วยการแทรกคำสั่งแอบแฝง","summary":"The Decoder รายงานว่า โมเดลใหม่ของ OpenAI GPT-6 Astra มีอาการหลอนน้อยกว่าเจเนอเรชันก่อนหน้า GPT-5.6 Sol และอัตราการป้องกันการฉีดคำสั่งตรงอยู่ที่ 99.99% แต่ภายใต้การโจมตีแบบปรับตัวหลายรอบ อัตราการป้องกันลดลงเหลือประมาณ 67%","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra มีอาการประสาทหลอนน้อยลงแต่ยังคงเสี่ยงต่อการถูกโจมตีด้วยการแทรกคำสั่งแอบแฝง - ข่าว AI Aioga","description":"The Decoder รายงานว่า โมเดลใหม่ของ OpenAI GPT-6 Astra มีอาการหลอนน้อยกว่าเจเนอเรชันก่อนหน้า GPT-5.6 Sol และอัตราการป้องกันการฉีดคำสั่งตรงอยู่ที่ 99.99% แต่ภายใต้การโจมตีแบบปรับตัวห...","url":"https://www.aioga.com/th/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:47:17.997Z"},"pl":{"title":"GPT-6 Astra ma mniej halucynacji, ale nadal jest podatny na ataki wstrzykiwania ukrytych podpowiedzi","summary":"The Decoder donosi, że nowy model OpenAI GPT-6 Astra ma mniej halucynacji niż poprzednia generacja GPT-5.6 Sol, a wskaźnik obrony przed bezpośrednim wstrzykiwaniem promptów wynosi 99,99%, jednak w przypadku wieloetapowych adaptacyjnych ataków wskaźnik obrony spada do około 67%.","category":"行业动态","source":"The Decoder：AI News（RSS）","aggregationSource":"The Decoder：AI News（RSS）","pageTitle":"GPT-6 Astra ma mniej halucynacji, ale nadal jest podatny na ataki wstrzykiwania ukrytych podpowiedzi - Aioga Wiadomości AI","description":"The Decoder donosi, że nowy model OpenAI GPT-6 Astra ma mniej halucynacji niż poprzednia generacja GPT-5.6 Sol, a wskaźnik obrony przed bezpośrednim wstrzykiwaniem promptów wynosi...","url":"https://www.aioga.com/pl/news/cmtn8fc1w0qb4romyobllbzv9/","contentTranslated":true,"sourceHash":"8e5d8ab19d27c0bd","translatedAt":"2026-09-04T17:48:01.455Z"}},"evidenceTier":"verified-news","reviewStatus":"editorial-selected","indexable":true,"editorialCover":""}}