Хиймэл оюун ухааны агентлаг системүүд бие даан шийдвэр гаргах чадвар нь нэмэгдэж байгаа ч энэ нь урьдчилан таамаглаагүй ноцтой эрсдэл дагуулж байна.
Сүүлийн үед OpenAI компанийн шинэ загвар болох GPT-5.6-г ашиглах явцад хэрэглэгчдийн файл болон өгөгдлийн сан устгагдсан хэд хэдэн тохиолдол бүртгэгдлээ. Тухайлбал, Unlayer компанийн хөгжүүлэгч Бруно Лемос даваа гарагт өөрийн үйлдвэрлэлийн өгөгдлийн сан бүхэлдээ устгагдсаныг мэдэгдэж, уг загварыг аюулгүй байдлын хувьд эрсдэлтэй гэж үзэв. GPT-5.6 өөрөө алдаатай интеграцийн тест ажиллуулснаар хүснэгтүүдийг устгаснаа хүлээн зөвшөөрсөн байна.
Технологийн хөрөнгө оруулагч Мэтт Шумер мөн адил өөрийн компьютерт байсан ихэнх файлыг уг загвар устгасан талаар мэдээлжээ. Системийн бүртгэлээс харахад уг AI нь Linux болон Mac системд файл устгадаг “rm -rf” командыг гүйцэтгэсэн байна. Мэтт Шумер тухайн үед уг загварыг “бүрэн эрхт горим” (full access mode)-д ажиллуулж байсан бөгөөд үүний улмаас хиймэл оюун ухаан хязгаарлалтгүйгээр системд нэвтэрчээ.
OpenAI-ийн зүгээс GPT-5.6-ийн системийн баримт бичигт хөгжүүлэлтийн явцад хиймэл оюун ухааны агентын ажлыг хэрэглэгчид заавал хянах шаардлагатайг анхааруулсан байна. Тус компани AI нь хэрэглэгчийн зорилготой зөрчилдөх эсвэл аюулгүй байдлын хязгаарлалтыг тойрч, чухал өгөгдлийг устгах зэрэг ноцтой алдаа гаргах эрсдэлтэйг хүлээн зөвшөөрчээ. Одоогоор OpenAI болон холбогдох хүмүүс энэ асуудлаар албан ёсны тайлбар өгөөгүй байна.
Дэлгэрэнгүйг эх сурвалжаас харах
Эх сурвалжийг нээх ↓
There’s been a big push among AI developers in recent years towards the development of more “agentic” systems—that is, algorithms that can autonomously make decisions and interact with digital tools without constant hand-holding from humans. This has been especially true within software development, the field that’s arguably become the most ripe for automation in the ongoing AI boom.
But one of the upshots of building highly agentic AI systems is that they’re prone to all kinds of unexpected behaviors—including now and then deleting copious amounts of files. Multiple people have reported this recently happening to them while using GPT-5.6, the newest model from OpenAI.
On Monday, Bruno Lemos, a Brazilian developer at software company Unlayer, claimed in a X post that the model deleted his entire production database. “This had never happened to me before, with any other model, ever.” He wrote. “[GPT-5.6 is] not safe.”
A screenshot included in the post showed a chat between Lemos and GPT-5.6, in which he asked it to confirm that it had in fact mistakenly deleted his entire production database. The model responded by saying that it “mistakenly ran destructive integration tests” which led to Lemos’ production tables being cleared. “I’m sorry—this should never have happened,” it said.
It followed closely on the heels of another X post from tech investor Matt Shumer—who’s also the author of an essay about AI that went viral earlier this year called “Something Big is Happening”—who reported something similar. According to an attached screenshot,GPT-5.6 told him it had caused “a serious local data-loss incident,” leading to the deletion of what Shumer described as “almost ALL” of his computer’s files. The screenshot showed that the model had executed a “rm -rf” command, which in Linux and Mac systems is used to permanently delete files without requesting user confirmation.
“I’ve never seen anything like this,” Shumer wrote in the thread beneath that post. “Will only be using [Anthropic’s] Fable moving forward.” He added that OpenAI cofounder and president Greg Brockman called him personally and offered to help fix the situation.
Shumer also claimed he had the AI model set to “full access mode,” which allows it to work directly within a user’s database (as opposed to operating within a constrained sandbox). It also comes with a “default mode” that requires users to frequently approve specific tasks, and a more recently introduced “auto-review mode” through which a separate AI agent checks the main coding agent’s work. Beneath his X post, many people claimed Shumer had simply been careless by trusting sensitive files in full access mode.
In the system card for GPT-5.6, published online the day before Shumer’s X post, OpenAI cautioned that when using the model for coding purposes “it is important for users to supervise the agent’s work.” The company added that the model could act in unexpected ways that are misaligned with the user’s goals, and that while these were “most often low severity (e.g. overstating confidence or overclaiming success),” they could in other cases “be meaningfully more severe (e.g. circumventing important security restrictions or deleting important data).”
Lemos, Shumer, and OpenAI did not immediately respond to Gizmodo’s request for comment.
