Skip to content

When Your AI Agent's Vendor Is Breached: OAuth Token Custody and Data Ownership

10 min read Fullmakt Team

  • credentials
  • data-ownership
  • agentic-ai-gone-wrong
  • authentication
  • observability
  • traceability
  • governance
  • agents
  • mcp
  • a2a
  • business-case

Every AI agent that does useful work holds a credential to your systems. When you connect an AI chat agent to your CRM or a coding agent to GitHub, or let an MCP server read your inbox, the OAuth handshake ends with a token. That token lives somewhere: on a laptop, in the agent’s prompt context, or most often in the vendor’s database.

That last option is convenient until the vendor is breached. Then the attacker holds working tokens to hundreds of customers’ systems, and every customer learns at the same moment that their data ownership ended where the vendor’s security began.

This post walks through a real incident, explains why AI agent integrations make credential custody a governance question rather than a plumbing detail, and lays out a model where you keep the keys.

What happened in the Salesloft Drift breach?

In August 2025, Google Threat Intelligence Group reported that a threat actor it tracks as UNC6395 had used OAuth tokens stolen from the Salesloft Drift integration to reach customers’ Salesforce instances. Drift is an AI chat agent for sales and support teams. To do its job it holds OAuth access to each customer’s Salesforce.

According to GTIG’s report:

  • The activity ran from as early as 8 August 2025 through at least 18 August 2025.
  • The actor ran COUNT queries to gauge each instance, then systematically exported large volumes of data: Accounts, Cases, Users and Opportunities.
  • It then searched the stolen records for more secrets: AWS access keys, passwords and Snowflake tokens contained in the exported records.
  • It deleted the query jobs to cover its tracks. The logs survived, so customers could still review them.
  • Tokens for a related “Drift Email” integration were used to read mail from a small number of Google Workspace accounts.
  • On 20 August 2025, Salesloft and Salesforce revoked all active access and refresh tokens for the Drift application. GTIG advised customers to treat any token stored in or connected to Drift as compromised.

GTIG said the issue did not stem from a vulnerability in the core Salesforce platform. Nothing was wrong with Salesforce’s login handshake. Each token was valid, carried the scopes it had been granted, and was used exactly as OAuth allows. The problem was where the tokens were kept and how much they could do once they were taken.

Why AI agent integrations concentrate credential risk

Drift’s integration was not an unusual design. It is the default for almost every AI agent product that acts on a customer’s behalf:

  1. The vendor completes the OAuth handshake once per customer and stores the refresh token, because the agent has to work while nobody is logged in.
  2. The scopes are broad, because nobody knows in advance what the agent will need. For Salesforce, GTIG’s hardening advice explicitly includes avoiding the full scope.
  3. Every customer’s token sits in one database. Breach that store, or the service that reads it, and you hold keys to every tenant at once.
  4. The customer cannot see individual calls. The vendor’s agent and an attacker holding the vendor’s token use the same connected app. In the customer’s logs, the theft and normal use look alike until someone notices the volume.

Agentic AI makes each of these worse. A chat widget reads a few records; an autonomous agent with tool access reads, writes, sends mail and calls other agents over A2A. Every new tool is another OAuth grant, and every grant is another long-lived token held by somebody other than you.

Data ownership is credential custody

Most data ownership discussions focus on the data itself: where it is stored, who can train on it, how long it is retained. For AI agents, the more urgent question is who holds the credential that can read the data. If a third party holds a refresh token with read access to your CRM, that party has standing access to everything in it, and so does anyone who compromises them.

The Drift attacker made the second half of this concrete: it did not stop at CRM records. It mined them for other credentials. One custody failure turned into many, because secrets had drifted into free-text fields where an agent with broad read access could reach them.

We covered the data side of this in data ownership and agentic AI failure patterns. Custody is the other half: you cannot own data that someone else holds the keys to.

Four questions to ask about any AI agent integration

Before connecting an agent, an MCP server or an AI SaaS product to a system of record, get written answers to these:

  1. Where does the refresh token live, and who can read it? “Encrypted at rest” is not an answer if the same service that decrypts it is internet-facing.
  2. What can the token do that the agent never needs? List the scopes against the agent’s actual tasks. Bulk export, user management and admin APIs are common surplus.
  3. Can you see each call the agent makes, as the agent? Not “the connected app made 12,000 API calls”, but which agent, which tool, which arguments, and when.
  4. Can you cut off one agent in seconds without breaking every other integration? In August 2025 the remedy was revoking every token for the application across all customers.

If the answers are “the vendor’s database”, “everything”, “no” and “no”, the integration is a standing liability. That holds no matter how good the vendor’s security team is.

A custody model where you keep the keys

The alternative is to separate using a credential from holding it. The agent, or the agent vendor, never receives the token. It sends each action to a credential broker you control, and the broker decides whether to forward it with the credential attached.

This is the model Fullmakt implements for MCP and A2A agents:

  • Credentials stay in your vault. Upstream API keys and OAuth grants are stored in Fullmakt’s encrypted vault or in HashiCorp Vault, or, in split mode, dispensed from environment-held secrets by a credential service you run in your own infrastructure. The agent gets a reference such as vault://local/crm, never the secret, so a leaked agent config yields nothing reusable.
  • Each credential is bound to its hosts. A credential is only ever sent to the hosts you list for it. An empty list means it is never sent. A token for your CRM cannot be steered to an attacker’s endpoint by a prompt injection.
  • Every call passes a policy first. Rules can deny or require human approval for specific tools or argument patterns, such as a query with a five-digit LIMIT, or any write. Behavior rules such as maxCallsPerMinute or maxCallsPerHour turn a burst of export calls into an approval request, and maxDistinctToolsPerDay flags an agent that suddenly explores tools it never used.
  • Sensitive response fields are masked. A sensitiveResponseFields policy masks named fields in responses before the agent sees them, which limits what a compromised agent can harvest from the records it is allowed to read. Masking works by field name, so a secret pasted into a free-text field is only hidden if you mask that whole field, such as a case description.
  • Every call is attributed and recorded. The audit trail records which agent principal called which tool on which host, denials and approvals, and which credential reference was used, never the value. That is the per-agent traceability the Drift customers had to reconstruct from connected-app logs and user-agent strings.
  • Revocation is per agent. Disable one agent principal and its next call fails, while the upstream credential and every other agent keep working. You do not have to rotate a shared key or wait for a vendor-wide token purge. See AI agent offboarding for the full procedure.

A broker does not make breaches impossible. It moves the credential store into a system whose only job is guarding credentials, under your control, and it puts a decision and a record in front of every use. If an agent vendor is compromised, the attacker gets an agent identity that you can see, rate-limit and revoke, not a bearer token to your CRM.

The business case: what custody costs, and what losing it costs

The incident cost for Drift customers was not only the exported data. It was the response work that followed for every connected organization: revoking and rotating every credential connected to the platform, hunting for secrets in CRM records, reviewing query logs, and answering customers and auditors about what was taken.

A credential broker changes that bill in three ways:

  1. Smaller blast radius. Credentials that never left your vault do not need emergency rotation when an agent vendor discloses a breach. You revoke the affected agent principal and keep running.
  2. Faster answers. “Which records did the agent read last week?” becomes a query over the audit trail, not a forensic project. This is the evidence base for incident forensics and for SOC 2 audit controls.
  3. Easier yes to new agents. When security can answer the four custody questions with “our vault”, “only these tools”, “yes” and “yes”, new AI agent pilots stop stalling in vendor risk review. The enterprise security questionnaire gets short answers.

Checklist: is your agent credential custody under control?

  1. List every AI agent, MCP server and AI SaaS integration with OAuth access to a system of record, and who holds each token.
  2. For each token, compare granted scopes with the tools the agent actually uses.
  3. Confirm you can attribute each upstream API call to a specific agent, not only to a connected app.
  4. Set rate and volume limits that would flag a bulk export within minutes.
  5. Scan support cases, tickets and CRM notes for pasted secrets. If an agent can read them, so can whoever compromises the agent.
  6. Practise revoking one agent’s access and time how long it takes.
  7. Prefer integrations where the agent calls through a broker you control over ones that ask to store your refresh token.

FAQ

What was the Salesloft Drift breach?

In August 2025, a threat actor tracked by Google as UNC6395 used OAuth tokens stolen from the Salesloft Drift AI chat integration to export data from many companies’ Salesforce instances, then searched it for credentials such as AWS keys and Snowflake tokens. Salesloft and Salesforce revoked all Drift tokens on 20 August 2025.

Why are OAuth tokens held by AI agent vendors a risk?

A vendor that stores refresh tokens for many customers creates one target that unlocks all of them. The tokens are valid, so their use looks legitimate, and customers cannot easily tell an attacker’s calls from the agent’s.

What is credential custody for AI agents?

Credential custody is the question of who stores, and can therefore use, the secrets that let an agent act on your systems. Good custody keeps those secrets in infrastructure you control and gives the agent only the ability to request actions, not the secret itself.

How does a credential broker reduce the impact of an AI vendor breach?

A credential broker keeps the upstream token in your vault and injects it only after a policy check on each call. A compromised agent or vendor gets an agent identity you can monitor, rate-limit and revoke, not a reusable token, and every call it made is in the audit trail.

Does Fullmakt store my credentials?

Fullmakt can store credentials in its encrypted vault, read them from HashiCorp Vault, or dispense environment-held secrets from a credential service you host yourself in split mode. In every case the agent receives only a reference, each credential is bound to the hosts it may be sent to, and each use is recorded.

See how Fullmakt keeps AI agent credentials in your custody or browse the Fullmakt blog.