By Hamza Faisal
The companies building the world's artificial intelligence are, in many cases, failing at the basics of cybersecurity. A new study of 18,160 AI companies worldwide found that more than one in three carries a critical unpatched vulnerability on its internet-facing systems — and the biggest names in the field fare even worse.
Black Kite, a third-party cyber risk intelligence firm, released its 2026 AI Supply Chain Report on November 10. Its research group assessed the externally observable security posture of AI vendors around the world, using September and October 2026 data. The headline finding: 34.2 percent of the companies studied had at least one critical-severity, publicly disclosed vulnerability that had not been patched.
Nearly one in ten — 9.8 percent, or 1,771 companies — carried at least one flaw listed in the US Cybersecurity and Infrastructure Security Agency's Known Exploited Vulnerabilities catalog, meaning attackers have already used the weakness in real-world attacks. Among the Forbes AI 50, 52 percent had critical unpatched vulnerabilities. Among large AI companies, the figure was 58 percent.
Basic hygiene is where the chain breaks
The report's most striking pattern is how ordinary the failures are. Weaknesses in email authentication — the checks that stop attackers impersonating a company — were found at 61.7 percent of AI companies for DMARC, 37.3 percent for DKIM and 11.2 percent for SPF. Corporate login credentials belonging to 6.8 percent of the companies studied turned up in records collected by information-stealing malware, rising to 19.9 percent among firms offering enterprise AI deployment solutions.
Almost three in ten companies — 29.7 percent — scored at or above the critical threshold on Black Kite's Ransomware Susceptibility Index, a measure of the externally visible conditions that put a company on ransomware groups' radar. In March 2026, companies in the highest susceptibility band suffered 291 times the attack rate of those in the lowest band, the report says.
One shared flaw, hundreds of victims
The study also shows how a single vulnerability can ripple across the whole AI ecosystem. Within a week of the disclosure of PHP2Shell, a pair of critical PHP flaws, Black Kite's scans found 939 AI companies — 5.2 percent of the study — running affected PHP versions on their internet-facing systems. The flaws, tracked as CVE-2026-109618 and CVE-2026-109619, were patched on October 21.
The report also revisits the March 2026 compromise of LiteLLM, the open-source tool that connects AI applications to language models, in which malicious code reached downstream environments. Black Kite identified 105 AI companies with security findings associated with LiteLLM and the workflow tools n8n and Langflow.
The newest risk: AI's own plumbing
The fastest-growing exposure may be the newest one. Black Kite examined 305 publicly discoverable Model Context Protocol server entries — the connectors that let AI applications reach business systems — attributed to 179 AI companies. Across those entries it recorded 1,732 security findings, of which 48.8 percent were rated High or Critical. Among the 77 Critical findings, 87 percent involved code or command injection.
Elsewhere, 44.3 percent of the AI companies assessed showed anomalous network activity, and botnet-related activity was recorded at 41.8 percent. The report's argument: buying AI creates more than a software relationship — suppliers often hold access to customer data and business infrastructure, so a supplier's weakness becomes the customer's exposure.
Black Kite stresses that its findings describe externally observed conditions, not confirmed compromises — the assessments were made from outside, without access to company systems. Even so, the report lands as a warning to every enterprise rushing to adopt AI: the vendors selling the future may not have patched the present.
