[KongchangAI]
· 2 min read· 1,349 words

US AI Secrets Are Leaking: Security Breaches Are Scarier Than Chinese Spies

US AI Secrets Are Leaking: Security Breaches Are Scarier Than Chinese Spies

Vox analysis: US AI giants push R&D under the banner of 'guarding against China' yet expose their own secrets due to weak basic security.

According to Vox, top US AI companies proclaim 'we can't let China win' while their core secrets are exposed by basic security vulnerabilities. Nvidia CEO Jensen Huang resigned over an alleged spy scandal, OpenAI had an employee install eavesdropping devices, and Anthropic's systems were breached in under 12 hours by former Doge employees. Vox compares their strategy to leaving the front door open while worrying about a window crack, noting they redact independent reports citing 'preventing China from getting secrets' while ignoring real weaknesses. The article argues AI companies must shore up physical security, cybersecurity, and internal controls before chasing trillion-dollar valuations and a $1.4 trillion IPO—otherwise weak security only helps rivals catch up.

Top US AI companies are facing an awkward reality: they loudly proclaim "we can't let China win," yet their own security is riddled with holes, exposing their most critical secrets to risk. According to an analysis by the YouTube channel Vox, a series of security incidents—from Nvidia to OpenAI to Anthropic—show that the real threat to these companies may not be external rivals, but lax internal management.

Frequent Security Incidents: From CEO Scandals to Conference Room Wiretaps

According to Vox, Nvidia CEO Jensen Huang was forced to resign due to an affair with a woman alleged to be a Chinese spy. If true, this news would undoubtedly deal a heavy blow to the company's reputation. But Vox points out this is just the tip of the iceberg.

But that's just the tip of the iceberg. OpenAI just had a spy of their own.

OpenAI also had its own "spy" problem. An employee allegedly installed eavesdropping devices in a conference room where sensitive conversations took place, attempting to obtain internal information. This physical-layer security breach exposes weaknesses in the company's employee background checks and physical access controls.

Anthropic's experience is equally concerning. According to Vox, a group of former Doge employees breached Anthropic's internal systems in under 12 hours, accessing the company's Slack chats and meeting transcripts. Completing an intrusion in such a short time shows that its internal system protections fall far short of the standard needed to safeguard core assets.

internal systems in under 12 hours, accessing the company's Slack and meeting transcripts.

These incidents collectively point to one fact: while US AI labs chase technological leadership, they have severely underinvested in basic security practices.

Doge (Department of Government Efficiency) is a temporary federal agency established by Trump after his inauguration in January 2025, led by Elon Musk, with the nominal goal of cutting government spending and streamlining federal agencies. During its operation, the agency gained broad access to IT systems and personnel data across multiple federal departments, sparking extensive controversy over data security and access boundaries. The "former Doge employees" mentioned in the Vox report refer to individuals who once worked at the agency and thus possessed certain system penetration or social engineering capabilities. When such personnel move to private AI companies, their access techniques and internal knowledge may constitute new security risks, which also explains why Anthropic's internal systems could be breached in such a short time.

The Logical Flaw in "Can't Lose to China"

In its analysis, Vox pointedly notes that labs like Anthropic often justify pushing AI development at full speed with the argument "we can't let China win," and use this as a reason to redact independent evaluation reports.

reports on their AI progress by saying, we can't let China glean any of our secrets.

However, Vox uses a vivid analogy: it's like worrying about a crack in the window while leaving the front door wide open. These companies may do some superficial work to prevent information leaks, but the real security holes—weak infrastructure, loose internal controls—could let secrets flow easily to rivals.

These companies may be practically giving away their secrets to China with their weak security.

In other words, if a company cuts corners on security, it ultimately cannot protect its lead. Instead, it only helps rivals catch up faster. Vox's logic is clear: true competitiveness comes not only from algorithms and compute power, but also from the basic ability to protect core assets.

Independent evaluation reports in the context of AI governance usually refer to public or semi-public documents produced after external third-party organizations (such as METR, Apollo Research, ARC Evals, etc.) conduct capability assessments, risk assessments, or red-team testing of frontier models. The value of such reports lies in providing an objective perspective unaffected by corporate PR departments, helping regulators, researchers, and the public judge the risk level of models. Redacting these reports on the grounds of "preventing China from obtaining secrets" actually undermines the effectiveness of external oversight: on one hand, the redacted parts are often precisely the most sensitive risk information; on the other hand, the credibility of independent evaluations depends on their completeness—once selectively edited, the credibility of their conclusions is also called into question. Vox's criticism is precisely that this approach substitutes a "national security" narrative for genuine security building.

Before Chasing Trillion-Dollar Valuations, Do Your Security Homework

At the end of its report, Vox raises a sharp question: if you claim to be building a "machine god" and, like Anthropic, are chasing a $1.4 trillion IPO, then you should at least get basic security right.

This question points to a deeper issue: when AI companies' valuations soar and capital pours in feverishly, security often becomes a cost item that gets sacrificed. But history has repeatedly proven that infrastructure that neglects security will eventually pay the price. For these companies wielding cutting-edge technology, a single serious leak could not only destroy trade secrets but also shake public trust and regulatory confidence.

From Nvidia to OpenAI to Anthropic, these cases are not isolated incidents but signals of systemic problems. If the US AI industry wants to maintain its technological leadership, it needs not only smarter models but also a more solid security culture.

Three Takeaways for the Industry

First, security investment is not optional. Whether physical security or cybersecurity, it must grow in tandem with R&D investment. Anthropic's internal systems being breached in 12 hours shows clear shortcomings in its security architecture.

Second, transparency and security are not contradictory. Redacting independent reports on the grounds of "preventing China from obtaining secrets" cannot replace genuine security building. On the contrary, open and transparent evaluation mechanisms help identify vulnerabilities and build trust.

Third, leadership behavior is security culture. From CEO-level scandals to eavesdropping by ordinary employees, the behavioral example set by senior management and the internal control atmosphere directly affect the entire company's security baseline.

Vox's report reveals an unsettling paradox: the companies most worried about technology leakage are precisely the most vulnerable on security. For AI giants chasing trillion-dollar valuations, doing their security homework may be more urgent than releasing the next model.

$1.4 trillion IPO refers to the listing valuation target discussed for Anthropic in the context of the report. For reference, this figure is close to or exceeds that of one of the few tech companies with the highest market capitalization globally today, far above the typical valuation range of AI unicorns in 2024. Such a high valuation expectation means the company needs to prove to investors before going public the sustainability of its business model and the controllability of risks. However, pre-IPO due diligence typically focuses on financial and legal risks, with relatively weak scrutiny of information security infrastructure. Vox's question is therefore practically relevant: if a company cannot even guarantee basic internal access controls and physical security before going public, then the "technological moat" implied in its valuation may be eroded by security shortcomings, and the risks ultimately borne by investors will be underestimated.

One-line Summary

Vox analysis points out that US AI giants push R&D under the banner of "guarding against China," yet expose their own secrets due to weak basic security.

Paragraph Summary

According to Vox, top US AI companies loudly proclaim "we can't let China win," yet their core secrets are exposed to risk due to rampant basic security vulnerabilities. Nvidia CEO Jensen Huang resigned over an alleged spy scandal, OpenAI was revealed to have an employee install eavesdropping devices in a conference room, and Anthropic's internal systems were breached by former Doge employees in under 12 hours, with Slack chats and meeting transcripts accessed. Vox uses the metaphor of "a crack in the window while the front door is wide open" to describe these companies' security strategies, noting they redact independent evaluation reports on the grounds of "preventing China from obtaining secrets" while ignoring real security shortcomings. The article argues that before chasing trillion-dollar valuations and a $1.4 trillion IPO, AI companies must shore up the basics of physical security, cybersecurity, and internal controls—otherwise weak security will only help rivals catch up faster.

#US AI security breach#Nvidia Jensen Huang#OpenAI spy#Anthropic hack#AI security vulnerabilities#Vox analysis
Share:

Related articles