At least nine AI developer platforms disclosed security incidents in October, and cloud security firm Wiz says the breaches share one worrying pattern: attackers hunting stored environment variables holding customers' AI provider keys. Wiz, which published its findings on October 30, has dubbed the run "Black October."

The nine platforms named by Wiz are Fireworks AI (October 6), Andromeda (October 9), Zeabur (October 14), Lightning AI (October 15), OpenCode (October 16-17), Railway (October 20), Warp (October 23), CommandCode (October 28) and Northflank (October 30). Wiz stresses it is not claiming a single actor behind all nine — only Railway has publicly used the word "campaign" — but every disclosure describes unauthorized access to credentials, most often environment variables carrying AI keys. Three platforms reported signs of an "agent swarm" attack; two blamed "distillation" attempts.

One pattern, many doors

The root causes differ from platform to platform. Fireworks, Zeabur and Lightning AI attributed their incidents to credential access; Railway believes the initial access came through privilege escalation in its build infrastructure; Warp pointed to a third-party vendor dependency; and OpenCode blamed a legacy system. Yet the target was the same in each case: the keys that let applications call frontier AI models.

That makes developer platforms an unusually rich target, since one platform can hold the provider keys of thousands of customers at once — and those keys are directly convertible into computing power. Only Zeabur has published a full write-up so far; it has said it will shut down because of the incident.

The scale of exposure varies. Railway says at most 1.85 percent of its workspaces, about 100,000, had environment variables potentially compromised, and has told customers to rotate any sensitive variable set before October 21. Northflank, the most recent platform on the list, told every customer to assume all environment variables had been decrypted and to rotate everything.

Stolen keys turned into millions of calls

Wiz traced two enterprise Amazon Bedrock keys from a single compromised vendor. The keys were first tested within 30 minutes of the reported credential access — a single model call from a residential internet address — then quietly re-checked for days before the flood began.

In one case, about 800,000 calls were made over roughly 30 hours, with about 90 percent succeeding. The attackers used cross-region inference profiles to dodge throttling, with only about 6 percent of calls blocked. In the second case, about 1.7 million calls were made in just 13 hours before regional quotas stopped the run. Together, the two cases totalled about 2.5 million calls, which Wiz values at more than $100,000 at list price.

Wiz links the key checks to two resale operations. One, which it calls "momo," is tied to an internet address hosting a site advertising discounted access to frontier models for cryptocurrency payment; the same address was publicly named by Zeabur as part of its compromise. The second, "Xingjiabi," runs a site offering discounted access to a wide range of frontier models. Wiz cautions that it cannot definitively tie either resale operation to the original breaches.

What developers should do now

Wiz's advice is blunt: treat any key stored on a platform that has disclosed an incident as compromised, and rotate it. The firm also recommends scoping every key to the narrowest model and region profile possible, and alerting on first use from an unfamiliar internet address — the pattern its researchers saw in both Bedrock abuse cases. The October run is a reminder that an AI key's security depends as much on the platforms storing it as on the code using it.