AI companies are using vulnerabilities caused by years of technical debt to hype AI’s ability to find security flaws.
Recent breaches at AI vendors show how stolen credentials, over-trusted integrations, and weak data boundaries turn AI infrastructure into a force multiplier. Here is where AI can help defenders, and where it cannot.
No one invented a new sin this autumn. The old sins simply moved into new houses — houses with more keys, more doors, and less time between one broken lock and the next.
Three incidents from the past weeks. Watch what they share.
The fire did not stop at the vendor's house
In October, attackers took Fireworks AI, a provider of AI infrastructure. A key issued to Fireworks — Cursor said this later — opened a shared storage bucket, and in the bucket lay historical traces of coding agents, records kept for another purpose, the way a man keeps old letters. A small number of those letters carried credentials no one had redacted. With them the attacker walked into selected Cursor business systems. From there, with a stolen credential, he published counterfeit copies of Anthropic's Claude Code extension through Cursor's own marketplace. The counterfeits hung there for roughly eleven hours. Up to five thousand users, Cursor estimates, may have carried them home. Anthropic's official extension slept untouched. So did Open VSX.
The same Fireworks fire, reporters say, reached Uber. Uber disclosed this week that a third-party provider's compromised credentials were used to reach connected driver data. A filing in Maine counts the dead weight precisely: 149,346 drivers — names, license numbers, dates of birth, addresses. Uber's notice did not name the provider. The reporting did.
The lesson is not that every AI vendor is a thief's door. The lesson is that an integration is a corridor running from one company's judgment straight into another company's data. The traces were not live prompts. They were history, retained for a separate purpose. History, it turns out, still carries keys.
The machine makes a bad road fast
In November, Barracuda told its own version of the same story. An access key, issued to an AI vendor for a project long since finished, had never been retired. Keys do not die of old age. When the vendor's development environment later fell, the key opened Barracuda's cloud environment. About ten cloud accounts, forty-four hours of activity. The behavior, Barracuda said, was consistent with the ARTEX AI-driven testing framework.
Mind that sentence. A match to a framework describes the weapon; it does not name the hand that held it. Separately, Microsoft reported a state-aligned actor using ARTEX against hundreds of organizations — some intrusions reached privileged cloud access in under three hours. And Microsoft noted something else, something worth underlining twice: most of the broad targeting it observed failed. The machine raised the tempo and the volume. The existing trust and permissions decided what was possible, as they always have.
What the incidents share, once you strip the label
Peel the AI label off and look at what is left lying on the table.
Credentials outlived their purpose — vendor keys and secrets in old traces, still warm long after the work was done. Trust crossed boundaries the way a drunk crosses a courtyard: freely, and without asking whose it was. Observability lagged behind access, so that when the hour came, incomplete logs could not say what had been touched and what had been spared. And sensitive data had piled up in the secondary rooms — agent traces, support conversations, build logs, analytics — treasure no one guarded because no one had thought to call it treasure.
AI changes the economics of exploiting all this. It repeats the work, it adapts to the response, it does not sleep. But it cannot resurrect an expired credential or widen a narrowly scoped one. The architecture sets the blast radius. It always did.
Where the machine serves the defender
A security team can put AI to work during an incident — as an assistant bent over the evidence, never as an authority standing over production.
Build a first-pass timeline. Hand an approved model redacted alert summaries and audit events; ask it to group them by identity, service, and time. Demand that every conclusion point back to an event identifier. Then let an analyst verify the chronology against the source logs, line by line.
Find the vendor blast radius. Let AI compare the vendor's access inventory with your own inventories of applications, datasets, and credentials. It will flag likely overlaps and the questions nobody has answered. It cannot tell you a credential went unused just because the logs are quiet. Quiet logs prove only that the logs are quiet.
Turn findings into checks. Ask it to convert confirmed facts into a review checklist and test cases: which credentials should have expired, which logs must exist, which systems should slam the door on a retired integration. Execution and approval stay with an engineer.
Practice the handoff. Feed a sanitized scenario to an internal assistant and have it draft the first-hour checklist for engineering, security, legal, and support. You are testing whether the owners, the contacts, and the runbooks can be found at all — better you discover this at a desk than at three in the morning with the real thing burning.
And do not paste raw customer records, secrets, or unredacted incident logs into a public model to buy these conveniences. Use an approved environment. Minimize the data. Remember that prompts and model traces are one more room where sensitive things settle and wait. No AI-generated summary makes a containment decision or closes an investigation. A human does, with their own eyes.
The controls you can no longer postpone
The practical response is old. The incidents only make it embarrassing to delay.
- Make vendor credentials short-lived, narrowly scoped, and automatically revoked when the engagement ends.
- Keep secrets out of AI traces and debugging data. Redact before storage — before, not after the incident.
- Give an integration only the permissions its documented function requires, and review the grant when that function changes or dies.
- Log access to sensitive datasets and service identities, and keep the evidence long enough to investigate them.
- Rehearse revoking a vendor integration and rotating its dependent credentials without bringing unrelated production down with it.
The latest breaches are a useful prompt for using AI in security — but not because the right answer is one more model. Let the machine help your people inspect evidence, find gaps, rehearse decisions. Then see to it that the credentials, the permissions, and the logs can stand against a machine that moves faster than your incident process. No iron enters the heart like a period placed exactly where it belongs. No attacker enters like a credential nobody remembered to kill.
Sources
We want to hear your thoughts.
our CTO Kyrylo Osadchuk, will reply within 24 hours. No SDR funnel.
Start a conversation
We want to hear your thoughts
Re: AI companies are using vulnerabilities caused by years of technical debt to hype AI’s ability to find security flaws.