A sales team vibe codes an AI agent to automatically update client records across the organization; a developer generates an API key to access customer data; or a developer forgets to close down a sandbox they were using to test an AI agent.
These three scenarios are common and the implications for data security are profound. Most of the security practitioners interviewed for this series agree that the challenge isn’t new. Cloud adoption and SaaS produced similar sprawl and data governance issues two decades ago. What's changed is speed and reach.
“This problem existed long before AI,” said Mandar Chaudhari, a full-stack geospatial engineer at Land IQ and a researcher at George Mason University. “Developers create test applications, temporary scripts, and database connections all the time. It gets more complicated when an AI tool is connected to a code repository, database or other application.”
For example: If a developer connects a coding agent to a test environment to help investigate an API issue and that environment has credentials with broader access than the task requires, then the agent might use those permissions to move laterally through the organization’s data stores. If that agent is left unmanaged or is abandoned and an attacker finds it, the attacker would have a sanctioned backdoor into all of the data and those systems, no hacking required.
And a lot of this access and data movement is not stopped by traditional access management tools such as privilege access management (PAM), data loss prevention (DLP), or next-generation firewalls (NGFW) because these tools will see an API as an authenticated non-human identity.
“I wouldn't say API keys automatically bypass DLP or firewalls,” said Chaudhari. “It depends on how those security systems are configured and what they're monitoring. The challenge is that an API request using valid credentials can look like normal application activity.”
Beware unmanaged third-party integrations
SaaS-to-SaaS agents are one of the hardest unsanctioned integrations to find and control because they work entirely in the background. Ryan Reck, founder of Reck Advisory, an independent cybersecurity and AI advisory firm, said AI tools can request broad access to mailboxes, shared drives and CRMs. Where self-service consent is allowed, employees can approve those connections without a procurement or IT review.
AI projects can also leave behind indexes, test datasets and logs that fall outside existing data governance structures. “Much of it is familiar from shadow IT and cloud sprawl,” Reck said. “AI adds more opportunities to connect systems and create copies of enterprise data.”
The risk isn't hypothetical. An April data breach at Vercel began when Context.ai, a third-party AI tool, was compromised. It led to the takeover of an employee's Google Workspace account, which the attackers then used to access that employee’s Vercel account and move laterally through the network, encrypting data as they went.
Unlike cloud developer integrations, which are one-to-one and set up individually by the developer, AI integrations often connect to several systems at once or to a model context protocol (MPC) server and can search, summarize, or move information between them, so one forgotten integration can expose data across multiple repositories.
Developer sandboxes are a perennial security issue
Forgotten developer sandboxes are a persistent problem. Cloud-based development platforms make it easy for developers to copy data into places that IT and security teams don’t know about, said Mudita Khurana, security engineer at Airbnb. AI exacerbates the problem because developers can connect third-party LLMs and agents directly to data repositories using API keys.
“One risk is intentional data being sent to an unapproved AI tool, but an even bigger risk is the unintentional side,” Khurana said. “In many cases, there could be instances where an integration was approved for a short experiment, but its access was never removed after the experiment ended.”
Because many API keys do not expire, the integration could still have a valid API key and access to source code and customer information long after the project is abandoned. If the key is exposed, attackers could use it to gain access without having to hack into the company first.
AI access points: forgotten API keys, OAuth consents and agent tokens
While the circumstances, players and technologies differ, most of the issues with AI gaining uncontrolled access to data revolve around credentials and delegated grants such as OAuth consents, forgotten API keys and service accounts, and agent tokens, said Bapuji Koraganti, a Principal Engineer at Comcast and independent identity researcher.
“The pattern isn't entirely new,” he said. “Cloud-era shadow IT was [also] about unmanaged assets, but the AI version is different. It's authorized but unaccountable access. Someone clicks ‘Allow’ once, and an autonomous process keeps using that grant with broad read scopes indefinitely.”
There are two risks that deserve more attention than they get, Koraganti said.
One is the revocation gap. This is when a leaked key or risky AI integration is revoked, but there are agent and integration layers that cache their own tokens or access refresh tokens that automatically renew access to system resources behind the scenes. Data may also have been copied into an embedding index that revocation never touches.
Another is permission flattening: This is where AI and retrieval-augmented generation (RAG) pipelines typically ingest data through one broadly privileged service account. And because source system row-level and document-level permissions don't carry into the vector store, anyone who can query the assistant can effectively reach data they could never open directly.
While these challenges are not new, the speed at which AI is being developed and deployed is outpacing the ability of most cybersecurity and data governance teams to keep up.
“A back door into company data can start at the front door, with an employee clicking ‘Allow’,” Reck said. “Someone needs to own that connection long after the trial ends.”