One In Three AI Suppliers Flagged For Critical Flaws. Black Kite Calls Them “Potential Exposure.”

Estimated reading time: 9 minutes

Buying AI means inheriting someone else’s supply chain. Black Kite scanned 18,160 AI suppliers and flagged a critical unpatched flaw at one in three. Dr. Ferhat Dikbiyik, its chief research and intelligence officer, calls each flag “potential exposure, not a confirmed hole.” Some cyber insurers use that data to decide whether to underwrite.

A company that adopts an AI service hands another company its data, Black Kite’s 2026 AI Supply Chain Report says. It may hand over parts of its business processes too. Sometimes it hands over direct access to its systems.

The supplier also has its own supply chain. It runs models, uses open-source components, and connects to customers through integrations. The report notes, “A weakness anywhere in that chain can affect the customer.” This includes software and services the customer did not buy directly.

The report gives an example. In March, attackers released a compromised version of LiteLLM, an open-source AI tool, which was available for 40 minutes. AI data provider Mercor installed it and later reported a breach. Mercor’s customers did not choose that package themselves.

Black Kite searched for similar weaknesses. Its report, released on Nov. 10, scanned 18,160 AI suppliers from the outside. “The AI supply chain isn’t failing in exotic ways. It’s failing in ordinary ones,” Dikbiyik said.

Black Kite sells the ratings that scans like these produce. Several cyber insurers use its data in underwriting, Dikbiyik said. Some set a score threshold for whether to write a company. Others share periodic Black Kite reports with their insureds. He did not name them. Its security chief made the same point to CINI this month.

The report recommends ongoing monitoring of suppliers, which is the service Black Kite offers.

Dr. Ferhat Dikbiyik, Black Kite.

What The AI Supply Chain Scan Found

Most of these suppliers are small companies. Nearly half of those with a recorded headcount have ten or fewer employees. For all of them, AI is at the core of their product.

  • 34.2% had at least one critical unpatched vulnerability.
  • 9.8% had a flaw on the federal Known Exploited Vulnerabilities catalog.
  • 61.7% had weak email authentication, the controls that stop spoofed messages.
  • 6.8% had employee logins in records stolen by infostealer malware.
  • 29.7% crossed Black Kite’s critical threshold for ransomware susceptibility.

What A Version Match Can Show

The 34.2% figure is based on version matching. Black Kite checks which software version a system uses and compares it to known critical flaws. The report notes that it does not test whether the flaw can actually be exploited.

Dikbiyik names the limit himself. A vendor can backport a fix without changing the version number. Black Kite cannot see that from outside. It reads build numbers and other details where they show. Even so, a finding is “potential exposure, not a confirmed hole,” he said.

One flaw received closer attention. Rapid Reset is a denial-of-service flaw in HTTP/2, the second version of the web’s Hypertext Transfer Protocol. It only works when HTTP/2 is enabled. Black Kite identified it at 2,902 companies based on version, then tested directly for the feature.

  • 963 companies were confirmed exposed.
  • 1,059 had the feature switched off on every system checked.
  • 880 sat behind content delivery networks and could not be verified.
See also  UIB Partners With CyberCube To Boost Cyber Liability Insurance Analytics And Growth In LATAM And Asia
What happened when Black Kite tested one version-based flag. Source: Black Kite 2026 AI Supply Chain Report.

The report counts only the 963, a third of the first flags. Of the companies Black Kite could check, fewer than half were exposed. Dikbiyik called Rapid Reset “the stricter case.” He did not describe a similar test for the other findings.

Rapid Reset also affects the known-exploited figure, making up 54% of the 1,771 companies with a known exploited flaw. It causes services to go offline but does not expose data. Without Rapid Reset, the rate drops from 9.8% to 5.5%, or 1,001 companies, according to Dikbiyik.

Network Signals Are Leads For A Conversation

Two more measures rely on internet protocol (IP) addresses. Botnet findings compare a company’s addresses to reports of botnet activity. These findings appeared at 6.5% of the companies, or 1,176 in total.

Shared cloud and hosting addresses can cause a company to be flagged because of a neighbor’s activity. Dikbiyik said the figures exclude shared IP addresses.

Unusual network activity was found at 44.3% of companies, or 8,037 in total. Black Kite compares each company’s traffic to its own baseline and flags connections to high-risk or unknown addresses, spikes in outbound data, and traffic on unexpected ports.

“Both signals are leads for a conversation, not verdicts,” Dikbiyik said.

Stolen Logins Follow The Supplier’s Work

Critical flaws appeared in about one in three AI companies of every type. Stolen logins were not as evenly spread; they clustered at suppliers that work most deeply inside customer systems, according to the report.

Cloud solution providers had the highest rate at 19.9%. Data and labeling companies followed at 13.9%, while model and platform developers were lower at 4.1%.

Company size makes a bigger difference. Infostealer records were found at 1.7% of the 8,612 companies with ten or fewer employees. The rate rises to 47.3% for companies with 251 to 1,000 employees, and 60.0% for those above that.

The report explains that the increase is not just because there is more to scan. Larger companies have more employees whose computers could be infected. Black Kite has not analyzed the data per employee, according to Dikbiyik. “It’s a fair point that bigger companies have more machines to infect.”

The records date from the last 90 days. Dikbiyik noted that the National Institute of Standards and Technology no longer recommends forced periodic password changes. He pointed to research showing stolen logins can stay usable long after the theft.

This is why Black Kite checks if the mailbox linked to a stolen login still receives mail. He did not say how many logins are still active. “A stolen credential is a question to put to the supplier, not proof of a breach,” he said.

One AI Component, Many Suppliers

AI products often share components. Black Kite found a known exploited flaw in n8n, a workflow tool, at 93 AI companies. Tools like n8n store the credentials that connect AI to other systems. According to CINI, Censys found that three out of four exposed n8n servers were running a release with a critical flaw.

See also  UK Enterprises Turn to Software-Based Pentesting Amid Rising Cyber Threats

The exposure goes beyond just AI companies. Black Kite’s customers monitor about 350,000 companies. Dikbiyik said that Black Kite first detected potential n8n exposure at 5,303 companies, and 1,163 still show it. For LiteLLM, the number dropped from 4,245 to 351.

These are potential exposures, not confirmed incidents. Dikbiyik called one component under many suppliers a concentration risk. Black Kite can watch it “shrink as suppliers patch,” he said. For insurers, shared parts mean accumulation risk. Suppliers that look independent can fail together.

Patch speed shows from outside too, though only in part. Black Kite flagged PHP versions affected by PHP2Shell at 939 AI companies the week after the Oct. 21 fixes. By Nov. 11, 656 still showed affected versions, Dikbiyik said. About 283 had moved off them in three weeks.

Some policies put a clock on that work. Chubb’s endorsement gives policyholders 45 days to patch a published flaw, CINI reported. Counted from the Oct. 21 fixes, that window would close Dec. 5.

The 656 is the maximum possible number. Dikbiyik said that Ubuntu, Red Hat, and the Surý repositories backported the fix without changing the PHP version, so a patched server can still appear vulnerable from the outside. He called the 283 a minimum for patching progress.

Get The Cyber Insurance News Upload Delivered
Subscribe to our weekly newsletter!

Ransomware Susceptibility: What The Index Counts

Black Kite’s Ransomware Susceptibility Index (RSI) ranges from 0 to 1. A score of 0.4 is the critical threshold, where companies become much more visible to ransomware groups, according to Dikbiyik.

About 29.7% of AI companies scored 0.4 or higher. In China, the rate was 48.5%, while in the United States it was 24.3%.

Black Kite tested the index using data from victims between April 2025 and March 2026. For companies scoring above 0.8, 41% were attacked. In the 0.4 to 0.6 range, 5.1% were attacked. For those below 0.2, the rate was 0.14%.

Most companies in the critical range were not attacked. Dikbiyik said that about 62,000 companies scored between 0.4 and 0.6 in a snapshot of Black Kite’s monitored base. More than 3,100 were attacked, while about 59,000 were not.

The index “doesn’t tell you who will be hit,” he said. It measures how visible and susceptible a company looks to ransomware groups. The report calls it a tool for prioritizing review.

Dikbiyik defends the rate. “In my view, a rare event that becomes 36 times more likely is not a small number.” The comparison is with the 0.14% band.

The count also misses some victims. Black Kite monitors about 190 ransomware and extortion groups, their leak sites, and public disclosures by victims. “If a company paid, never disclosed, and never appeared on a leak site,” he said, “it isn’t in our count.” He believes the real rates are “if anything, higher.”

See also  SecondSight Exits Stealth Mode with $3 million Seed Round; Applies AI to Digital Risk Analysis for Cyber Insurance

What No Outside Scan Shows

The report’s case study shows the limits of outside-in data. In 2025, attackers stole access tokens for Drift, an AI chat tool from Salesloft, and used them to access data from customers’ Salesforce systems. Google estimated that more than 700 companies may have been targeted.

No external scan of Salesloft would have shown Drift’s standing access to that data, the report says. Insurers cannot yet see that kind of aggregation, Nudge Security’s Russell Spitler told CINI.

Dikbiyik leaves these questions to underwriters, saying, “I’m not an underwriter, so I won’t tell underwriters what to ask.”

He does identify the blind spot: it is inside the customer’s own environment. No outside scan can see which systems an AI tool can access, what permissions it has, or how quickly access can be revoked.

“In my view, the access that does the most damage is the access you granted,” Dikbiyik said. “Outside evidence then tells you which suppliers to talk to first.”

FAQ – AI Supply Chain

What is AI supply chain risk?

The cyber exposure a company inherits from its AI suppliers and from the models, tools and integrations those suppliers depend on.

How many AI suppliers have critical vulnerabilities?

Black Kite flagged at least one at 34.2% of 18,160 AI companies. It says version-based findings show potential exposure, not confirmed holes.

How reliable are version-based flags?

Black Kite tested one flaw, Rapid Reset. Of 2,902 companies flagged, it confirmed 963. Some vendors fix software without changing the version, which a scan cannot see.

How does Black Kite count botnet findings?

It links a company’s internet protocol (IP) addresses to sources reporting botnet activity. Findings appear at 6.5% of companies. Black Kite says shared IP addresses are excluded.

Which AI suppliers have the most stolen logins?

Cloud solution providers, at 19.9%, and data and labeling companies, at 13.9%. Model developers were among the lowest at 4.1%.

Why does the AI supply chain matter to insurers?

Many suppliers share the same components. One compromised tool can reach several vendors and their customers at once.

1 thought on “One In Three AI Suppliers Flagged For Critical Flaws. Black Kite Calls Them “Potential Exposure.””

Leave a Comment

×