More than half of the promising AI companies developing and providing ultra-high-performance artificial intelligence technologies have been found to have critical security vulnerabilities. An investigation targeting approximately 18,000 AI companies worldwide revealed that one in three has severe security flaws, heightening concerns over the overall security management of the AI industry.

According to a recent evaluation of the external security posture of 18,160 AI companies globally by the U.S. security firm Black Kite, 26 out of 50 promising AI companies selected by Forbes (52%) were found to have critical, unpatched vulnerabilities.

Black Kite reported that at least one publicly disclosed vulnerability of critical severity was identified in 34.2% of the surveyed AI companies.

Among them, 1,771 companies (9.8%) possessed at least one vulnerability included in the Known Exploited Vulnerabilities (KEV) catalog maintained by the U.S. Cybersecurity and Infrastructure Security Agency (CISA), which lists flaws known to have been actively exploited in real-world attacks.

The threat of ransomware attacks—which encrypt data stored on servers to block access—was also confirmed. Based on Black Kite’s Ransomware Susceptibility Index (RSI), which measures susceptibility to ransomware, 29.7% of all AI companies scored 0.4 or higher, marking the threshold for severe risk. Among these, 11.5% fell into the high-risk zone with an RSI of 0.6 or higher.

Abnormal security signs were also observed in some companies. Unusual network communication activities were detected in 44.3% of the surveyed AI companies. Separately, 6.5% of the companies held IP addresses reported to be associated with botnet activity. Botnets are collections of servers infected with malicious code and remotely controlled by hackers. Furthermore, evidence that corporate account information had been harvested by infostealer malware was identified in 14 companies (28%).

Ferhat Dikbiyik, Chief Research and Information Officer (CRIO) at Black Kite, pointed out that companies developing highly sophisticated AI technologies are failing to patch critical security vulnerabilities and are exposing credentials that provide access to sensitive systems, revealing a lack of basic security management.

The greater concern is that the security vulnerabilities of these AI companies do not just end as isolated damages for individual firms, but can serve as attack pathways into their clients' systems.

"The more access rights businesses grant to AI suppliers regarding their data, applications, and infrastructure, the more they inherit the security vulnerabilities of those suppliers," CRIO Dikbiyik stated.

This highlights how vulnerabilities in AI suppliers can spread along software and service supply chains, given the broad access rights companies typically grant to them.

A recent corporate account information leak incident at Anthropic demonstrates how security vulnerabilities in external software can threaten the internal systems of an AI company.

On the 31st of last month, Anthropic reported to the Personal Information Protection Commission that information from 112 domestic corporate accounts using its AI service, Claude, had been leaked.

The leak is estimated to have occurred when an external security firm gained unauthorized access to Anthropic’s internal systems. Previously, The Washington Post (WP) reported on the 26th of last month (local time) that Cathedral had exploited vulnerabilities in Anthropic's enterprise software to access internal systems.

Cathedral explained that it accessed Anthropic's systems to expose the security vulnerabilities of major U.S. AI companies. According to a draft report by Cathedral, an AI agent managed to find pathways to sensitive internal data—such as meeting minutes, payroll and hiring information, and Slack messages regarding advanced AI models—inside Anthropic within just 12 hours.

Anthropic stated that this incident may have exposed account and organization IDs, subscription tiers, and service usage and expenditure metrics for Claude enterprise account users.

The Personal Information Protection Commission officially launched an investigation into the incident on the 10th.

[Copyright (c) Global Economic Times. All Rights Reserved.]