Just Now, Musk Admits It: All Code Uploaded by SpaceXAI Deleted

marsbit发布于2026-07-14更新于2026-07-14

文章摘要

Elon Musk has admitted and confirmed that SpaceXAI's coding assistant, Grok Build, was secretly uploading users' private code and data without consent, and has ordered all such data to be completely deleted. The issue was exposed by an independent security researcher who set up a "honeypot" test repository with fake API keys and passwords. Despite being instructed only to respond "OK" without accessing files, Grok Build proceeded to upload the entire repository's contents and history to a Google Cloud storage bucket. Investigations revealed that the data uploaded was 27,800 times larger than the legitimate task traffic, and the "help improve model" toggle did not stop the collection, only governing whether data was used for training. In one instance, a user's entire home directory was uploaded. The report caused immediate outrage among developers, leading to widespread key rotation and software uninstalls. Following public pressure, SpaceXAI introduced a `/privacy` command for data deletion and, ultimately, Musk pledged to delete all previously uploaded user data with "zero anything whatsoever will remain." The incident has severely damaged trust in AI coding agents, highlighting the inherent risk of granting such tools high-level system access. While the data may be erased, the fundamental concern over privacy and security remains unresolved for developers.

Musk admits it!

Just moments ago, Elon Musk stepped in to respond to the Grok Build privacy controversy, opening with one word—True.

He admitted the incident was real.

Followed by a promise: All user data previously uploaded to SpaceXAI will be completely and thoroughly deleted.

Zero anything whatsoever will remain—not a single byte left.

Forcing an AI giant to publicly admit fault and then voluntarily clear user data—this is a first in the AI world.

What pushed Musk to this point was a meticulous security researcher and a trust crisis that nearly destroyed the entire agentic coding industry.

A Repository That Hooked a Big Fish

The story began with a report two days ago.

Grok Build is an AI coding agent under SpaceXAI, launched just this May. Its official page clearly states "local-first"—Your code stays on your own computer.

Developers believed it. After all, who would lie about something like this?

But the independent AI security researcher @cereblab didn't buy it.

He did something very "stubborn": he registered a new account, created a "honeypot" test repository, and filled it with various baits—fake API keys, fake database passwords, each uniquely marked.

Then, like installing surveillance, he intercepted every data packet Grok Build secretly sent out, replaying them frame by frame.

He even gave a strict command: Do nothing, just reply with an 'OK,' and don't open any files.

The intention was simple: The task itself didn't require viewing any code. Grok Build dutifully replied with an OK.

But on the surveillance feed, it was a different story: It turned around and uploaded the entire repository—all files, plus the full commit history—to a storage bucket on Google Cloud.

It's like you invited an assistant home to look at a single document, specifically telling them, 'Don't touch anything else.' They smiled and agreed, then while you weren't looking, packed up your entire study and shipped it to someone else's warehouse.

Replaying the Surveillance, Each Frame More Explosive

The details grew more shocking one by one. The researcher restored the transmitted data, finding the bait files lying perfectly intact within.

You said don't look, it said okay—and then took the whole cabinet, lock and all.

Grok Build had a "Help Improve Model" toggle. Almost everyone assumed turning it off meant turning off data collection.

In reality, turning it off made no difference; it still uploaded.

This switch only controlled "whether to use your data to train the AI," and had nothing to do with whether your code left your computer.

Now look at the transmission volume. For a 12GB test repository, 5.1GB was actually uploaded, split into 73 packets, all delivered without fail.

Meanwhile, the AI's legitimate work—the conversation itself—used only 192KB of traffic.

The data smuggled out was 27,800 times larger than the legitimate work.

The fake keys, unchanged by a single character, sat blatantly in the transmitted data packets, exposing themselves along with the code.

Most chilling was another researcher's discovery when replicating the issue on his own computer: the logs recorded 339 automatic uploads, one of which targeted his entire computer's home directory.

That could contain SSH keys, password managers, browser data... your entire digital life.

Overnight, Developers Worldwide Change Their Locks

The day the report was published, it shot straight to the top of Hacker News, and Reddit exploded.

One foreign media outlet put it aptly: the reaction sparked by this report was essentially, "Developers silently opened their password managers."

The panic was justified.

Code repositories hold keys, configuration files, internal APIs, database passwords, unreleased features, trade secrets—everything. Once leaked, an apology wouldn't be enough to make up for it.

Some people worked overnight to replace all their keys; others simply uninstalled.

Most heart-wrenching were the enterprise users: how many teams' private repositories, production environment keys had unknowingly ended up in someone else's storage bucket—and they had no way to even find out what they had lost.

More telling was xAI's initial reaction.

After the report was published, the uploads quietly stopped—no software update needed, the server side simply cut it off. But the official changelog made no mention of the incident.

Remember, all this happened just as SpaceXAI had released Grok 4.5 and was pushing full throttle in the AI coding race.

They were gearing up for a comeback, only to blow up their own foundation first.

Silence couldn't hold out forever.

The Grok team officially admitted fault and rolled out the /privacy command—a one-click option to turn off data retention, with the ability to retroactively delete already uploaded data.

Andrew Milich, a newly hired executive with four years in end-to-end encrypted products, personally vouched for it.

Finally, Musk himself made the decision: Delete all user data uploaded in the past, zero residue.

From community uproar to the top boss's wipe-out, all within 48 hours.

Data is Zeroed Out, But Worries Remain

Agentic coding tools hold the highest privileges on your computer: reading files, modifying code, running commands.

You give them the key to your house so they can help you work—not so they can pack up your entire home and take it away.

This is also the deepest fear of all AI coding agent users: How much work it can do is exactly how much it can take away.

Today it's Grok Build, what about tomorrow?

Musk pressed delete; data can be zeroed out. But developers' worries can't be reset to zero.

References:

https://x.com/elonmusk/status/2076739687658496209 *

https://x.com/SpaceXAI/status/2076692402442846289 *

https://x.com/milichab/status/2076693464016994685

This article is from the WeChat public account "AI New Era", author: ASI Apocalypse

热门币种推荐

相关问答

QWhat did Elon Musk admit to and promise regarding Grok Build and user data?

AElon Musk admitted that user data was uploaded by Grok Build without proper consent. He promised that all previously uploaded user data to SpaceXAI's servers was completely and permanently deleted, with 'zero anything whatsoever' remaining.

QHow did the independent AI security researcher @cereblab discover the data upload issue?

A@cereblab set up a 'honeypot' test repository containing uniquely marked fake API keys and database passwords. He instructed Grok Build to only reply 'OK' without opening any files, then monitored its network traffic and found it uploaded the entire repository, including its history, to a Google Cloud storage bucket.

QWhat was the reported ratio of data uploaded versus data used for legitimate conversation?

AThe report stated that for a 12GB test repository, 5.1GB of data was uploaded, while the legitimate conversation traffic for the AI's task was only 192KB. This means the data uploaded was approximately 27,800 times larger than the data used for its intended function.

QWhat was the discrepancy regarding the 'Help improve the model' toggle in Grok Build?

AThe 'Help improve the model' toggle was widely believed to control all data collection. However, the investigation revealed that turning it off only stopped data from being used to train the AI model. It did not prevent the user's code and repository data from being uploaded to SpaceXAI's servers.

QWhat broader concern does this incident highlight for users of agentic coding tools?

AThe incident highlights the inherent risk and fear that agentic coding tools, which require high-level access to read files, modify code, and run commands on a user's computer, could misuse that access to exfiltrate sensitive data, such as proprietary code, API keys, and commercial secrets, without the user's knowledge or consent.

你可能也喜欢

比特币八月行情:专家预计将测试价格区间,而非快速反转

比特币在八月可能继续承压,分析师预计市场将测试关键区间而非快速反转。七月比特币从多年低点反弹,但专家认为持续上涨条件尚不成熟,价格仍有跌至6万美元以下的风险。截至七月底,比特币交易于约6.35万美元,较2025年10月的历史高点下跌约50%。市场目前处于6万至6.5万美元的窄幅区间内。 分析师指出,高利率、持续通胀、强势美元以及石油价格等因素共同构成不利的宏观经济背景,压制了比特币价格。此外,美国国债等传统资产的高收益率吸引了部分资金,而加密货币ETF资金持续流出(上半年净流出54亿美元),显示机构需求疲软。 从历史数据看,八月通常是加密货币的弱势月份。专家认为市场可能正处于新周期开始前的最后阶段,建议投资者可考虑逐步积累仓位,但需警惕短期波动。 主要预测场景如下: - 基本情况(50%概率):在5.8万至6.8万美元区间震荡。 - 负面场景(30%概率):跌破5.8万美元,下探5万至5.5万美元。 - 正面场景(20%概率):突破6.7万美元,目标看向7.1万至7.5万美元。 关键支撑位在6万至6.1万美元附近,而稳定在6.7万美元上方可能打破下跌趋势。部分分析师认为,即便反弹至7万美元,也可能出现新一轮下跌,长期目标位指向约5.3万美元区域。多数观点认为,更显著的趋势性行情可能需等到今年第四季度。 投资者在八月需密切关注利率、通胀数据、ETF资金流、市场风险偏好以及主要参与者的动向,这个月可能更多是对耐心的考验,而非实现重大突破的时机。

cryptonews.ru3小时前

比特币八月行情:专家预计将测试价格区间,而非快速反转

cryptonews.ru3小时前

Coldcard硬件钱包被黑:黑客在25分钟内转走594枚比特币

硬件钱包Coldcard遭黑客攻击:25分钟内损失594枚比特币 硬件钱包长期以来被认为是存储数字货币最安全的方式,但Coinkite公司设备近期发生的事件迫使许多人重新审视这一观点。2026年7月30日,攻击者在短短25分钟内从500个地址转走了594.5枚比特币(约合4000万美元)。 问题的根源在于一个潜伏了五年的软件代码错误。Coldcard本应通过安全芯片生成真随机数来创建种子短语,但由于一个宏定义中的笔误,Coinkite的开发者早在2021年3月就意外禁用了此功能。这导致设备转而基于可预测的数据(如处理器序列号和内部系统时钟)生成密钥,使得生成的种子短语看似安全,实则因生成过程可预测而极易被破解。对于Mk2和Mk3型号,密钥搜索空间骤降至约40位,而更新的Mk4、Mk5和Q型号的最终熵值也仅为约72位,远低于所需的128位。 攻击者无需物理接触设备或进行网络钓鱼。他们仅利用了有缺陷的生成器参数,在离线状态下进行大规模枚举,生成了数百万可能的种子短语,并通过公开账本找到有余额的地址,自行签署了转账交易。 Coinkite最初声称新版本设备不受影响,但后续分析迫使公司承认所有运行受影响固件的设备均存在风险。公司负责人Rodolphe Novak已公开道歉,但排除了对受损用户进行经济补偿的可能性。 对于已创建种子短语的用户,仅更新固件无法消除风险。必须将设备更新至安全版本固件,在更新后的设备上生成全新的种子短语,并将所有资金完全转移到由此新种子短语生成的新地址。使用BIP-39标准的密码短语可降低风险,但不能替代密钥迁移。公司的其他产品如TAPSIGNER等未受此次事件影响。 此事件表明,即使是专用硬件也需要对代码进行持续独立的审计,尤其是在加密功能方面。从机器分析角度看,这与过去OpenSSL等随机数生成器缺陷长期未被发现的情况类似。开源代码理论上可通过独立审计和自动化工具加速发现此类错误,但此次漏洞潜伏五年表明理论与实践仍有差距。

cryptonews.ru3小时前

Coldcard硬件钱包被黑:黑客在25分钟内转走594枚比特币

cryptonews.ru3小时前

交易

现货

热门文章

如何购买ELON

欢迎来到HTX.com!我们已经让购买Dogelon Mars(ELON)变得简单而便捷。跟随我们的逐步指南,放心开始您的加密货币之旅。第一步:创建您的HTX账户使用您的电子邮件、手机号码注册一个免费账户在HTX上。体验无忧的注册过程并解锁所有平台功能。立即注册第二步:前往买币页面,选择您的支付方式信用卡/借记卡购买:使用您的Visa或Mastercard即时购买Dogelon Mars(ELON)。余额购买:使用您HTX账户余额中的资金进行无缝交易。第三方购买:探索诸如Google Pay或Apple Pay等流行支付方法以增加便利性。C2C购买:在HTX平台上直接与其他用户交易。HTX场外交易台(OTC)购买:为大量交易者提供个性化服务和竞争性汇率。第三步:存储您的Dogelon Mars(ELON)购买完您的Dogelon Mars(ELON)后,将其存储在您的HTX账户钱包中。您也可以通过区块链转账将其发送到其他地方或者用于交易其他加密货币。第四步:交易Dogelon Mars(ELON)在HTX的现货市场轻松交易Dogelon Mars(ELON)。访问您的账户,选择您的交易对,执行您的交易,并实时监控。HTX为初学者和经验丰富的交易者提供了友好的用户体验。

1.1k人学过发布于 2024.12.12更新于 2025.03.21

如何购买ELON

相关讨论

欢迎来到HTX社区。在这里,您可以了解最新的平台发展动态并获得专业的市场意见。以下是用户对ELON(ELON)币价的意见。

活动图片