FENCE · Worked case
Salesloft Drift, 2025
Stolen vendor tokens were used to export customers' Salesforce data
In August 2025 an intruder used stolen OAuth tokens of Salesloft's Drift chat agent integration to pull Salesforce data from many companies, Cloudflare among them. We read the public record against FENCE's five steps: where each check meets the attack path, and where none does.
- Scope
- One customer's side: Cloudflare
- Read against
- FENCE 1.1.1Requirements unchanged in 1.1.2
- Evidence
- 14 public claims citedEvidence frozen
- Reading
- One reading, with hindsightNot independently repeated
38 requirements reviewed
- 17
- are on the reported attack path
- 9
- apply, but are off that path
- 12
- do not apply to this setup
None is marked Pass or Fail. The public reports say where the attack went, not how any customer’s controls stood before it, so the 26 requirements that apply stay Unknown.
On the attack path, by step
The 12 that do not apply: all five Enclose requirements, three under Name, one under Constrain, and the three for payments.
On our reading, stolen tokens, not the Drift agent, carried the attack from Salesloft's GitHub into customers' Salesforce, where 17 FENCE checks meet it.
Drawn from GTIG, Cloudflare and Salesloft write-ups
Swipe for the full figure →
Our reading: the tokens did the work, not the agent
From as early as August 8, 2025 through at least August 18, an intruder used stolen OAuth tokens of the Salesloft Drift app to export large volumes of data from many companies' Salesforce instances.1 Google Threat Intelligence Group (GTIG) assesses its main aim was credentials, such as Amazon Web Services (AWS) access keys and passwords.1 At Cloudflare what it took was support-case text holding 104 Cloudflare API tokens.2,3 On our reading the intruder used its own tools, not the agent. If so, a check limited to the agent's own tool calls would not have seen those requests.
If a vendor holds a token into your systems, this is the exposure. The intruder sat in Salesloft's GitHub account from March through June before reaching Drift's AWS environment; Salesloft dates the intrusion March 22 to September 5.4 Then the ending: Salesloft revoked Drift-to-Salesforce connections on August 20; Cloudflare had no sign the action concerned it, and was told on August 23.5
We assess with hindsight, from the write-ups of GTIG, Cloudflare and Salesloft and the Risky Business hosts' account of them.6 They say where the attack went, not how any victim's controls stood, so nothing passes or fails. Of FENCE 1.1.1's 38 live requirements, 26 are Unknown and 12 do not apply: the customer neither built nor hosted the agent, no Model Context Protocol (MCP) server or agent-to-agent messaging was involved, and no money moved.
The decision rule still applies. The integration's credential reached customer data,2 so FENCE rates it High impact, and at that level any Unknown on a requirement that applies blocks expansion. On the public record alone, then, nobody could show that this integration met FENCE's bar. Of the 26 Unknowns, 17 sit on the attack path; your own records could settle many of those.
What to check in your own systems
- Which vendors hold a token to your Salesforce, mail or cloud? Within two days of being told, could you disconnect every third-party integration and issue new secrets for each service, as Cloudflare did for Salesforce?5,7
- Do you know where each vendor's traffic should come from, and would its token used from elsewhere raise an alert?
- Of the 17 on-path checks the record leaves Unknown, how many could your own records settle today?
Find
Find them: an account, a secret, and no list we can see
Cloudflare could find the Drift integration and cut it off: by August 25 it had disabled the Drift user account, revoked its client ID and secrets, and disconnected every other third-party integration from Salesforce.7 The reports do not show what records existed before the breach: whether the account, its secrets and the other integrations were listed, with owners and a way to revoke them all.
Register every agent, including the ones employees built, with a named owner and what it can reach.
Drift had its own user account in Cloudflare's Salesforce, disabled by August 25.7 Whether that account was on a register with a named owner is not public, and Cloudflare does not say what access it held.2
Find the agents and MCP servers that run on a shared company-wide key or a static secret, and fix them first.
The integration had a client ID and secrets, revoked by August 25.7 Whether those secrets were static, listed and put first for replacement is not public.
List every third-party integration that holds OAuth tokens into your SaaS, what each can reach, and how you would revoke all of them at once.
Within two days of the August 23 notice, Cloudflare disconnected all third-party integrations from Salesforce and issued new secrets for every service.5,7 So it could find those integrations and disconnect them during its response. Whether the list, each integration's reach and a way to revoke them all were written down before the breach is not public.
For each agent or integration on your agent register (FENCE-FIND-001) or integration list (FENCE-FIND-003), list the credentials it holds into your systems that cannot be revoked in one place, such as API keys held in each provider, with the provider that holds each one and an owner who can revoke it there.
Salesloft revoked the OAuth tokens itself and left API keys to be revoked in each provider.8 Whether Cloudflare had connected Drift with any API keys of its own is not public.
2 more, off the attack path
Pick one inventory as the record when platforms keep their own lists of the same agents.
The reports say nothing about competing inventories.
Watch live activity as well as scheduled scans, so an agent that lives for minutes is still found.
Cloudflare reports activity from August 9 to August 17.9 No short-lived agent is described.
Enclose
Enclose it: nothing to enclose in the vendor's cloud
Drift reached Salesforce as a connected app.1 The reports describe no sandbox, gateway or MCP server on the customer's side and no messages between agents, so all five checks are Not applicable. Salesloft lists 20 addresses it calls confirmed malicious and three user-agent strings.10 We read those as the intruder's own tools, not the agent, using the stolen tokens.
All 5 not applicable
- FENCE-ENC-001
Not applicable
Run coding agents in a container or VM with limited outbound traffic, and keep SSH keys and long-lived secrets out of their reach.
Drift is a chat agent the vendor hosts, not a coding agent the customer runs.
- FENCE-ENC-002
Not applicable
Enforce a sandbox's network limits outside the agent, and keep checking that they hold.
Drift ran in the vendor's environment;4 the reports describe no sandbox run by the customer.
- FENCE-ENC-003
Not applicable
Treat third-party skills and MCP servers as untrusted code, and review their scripts and instructions before install.
The reports describe no skill or MCP server; the integration was a Salesforce connected app.1
- FENCE-ENC-004
Not applicable
Patch MCP SDKs and AI gateways like any internet-facing software, and check that each gateway's admin API requires authentication.
The reports describe no MCP SDK or AI gateway.
- FENCE-ENC-005
Not applicable
Treat messages from other agents as data, never as orders.
The reports describe no messages between agents.
Name
Name it: an account and a secret are not yet a named agent
Cloudflare reports a Drift user account with a client ID and secrets.7 Whether that amounted to an identity of the agent's own, with short-lived tokens, is not public. The larger question sits with the vendor: the tokens came from Drift's AWS environment, and neither Salesloft update says how the intruder got into GitHub or on to AWS.4
Give each agent its own workload identity, exchange it for short-lived tokens, and stop sharing static keys between agents.
Cloudflare reports a Drift user account, a client ID and secrets.7 Salesloft and Salesforce revoked all access and refresh tokens for the Drift app,11 but Cloudflare does not say what kind of token its integration used,2 and no report gives token lifetimes.
Keep credentials out of the model's reach. Agents get them from a gateway or the identity provider, and no secret sits on local disk.
The tokens were taken from Drift's AWS environment, and the intruder also stole environment-variable secrets.4 This is the vendor's central question: the reports do not say where in AWS the tokens sat or what guarded them.
3 more, not applicable
- FENCE-NAME-003
Not applicable
Where SPIFFE/SPIRE already runs, build agent identity on it.
No workload of the customer's own was involved; the agent reached Salesforce over OAuth from the vendor's side.
- FENCE-NAME-004
Not applicable
If you run Entra, Google Cloud or Okta, evaluate its agent identity feature before you build your own, including which agents and runtimes it covers, what it can revoke and how it models an agent.
The customer did not build this agent.
- FENCE-NAME-005
Not applicable
Give each cloud-hosted agent its own execution role, or another identity the platform assigns and the agent cannot set.
The agent was not hosted in the customer's cloud. For the vendor this check would be Unknown: Salesloft describes a review of its identity and access management roles after the breach,12 not the roles in place before it.
Constrain
Constrain it: three minutes was enough for the support-case text
At Cloudflare a Salesforce Bulk API 2.0 job ran from 11:11:56 to 11:15:18 UTC on August 17 and took the support-case text.9 Whether a chat agent needed all case text, or for how long, is not stated. The check on an agent's own tool calls is off the path on our reading: the requests came from the intruder's tools, not the agent.
Delegate, never impersonate. Use token exchange so the agent keeps its own identity and never holds the user's token, and have the authorization server limit each token it issues to the next service and the task.
Drift had its own user account at Cloudflare.7 That does not show whether the integration acted under its own identity or held a user's token. Whether each token was limited to the next service and the task is not public; the stolen tokens were used to query Account, Opportunity, User and Case objects.1
Keep access tokens to one hour or less, keep signing keys in a KMS or HSM, check your authorization server against RFC 9700, and keep each token narrow in audience and scope.
Salesloft and Salesforce revoked access and refresh tokens for Drift;11 no report gives the lifetime of either. With the stolen credential, one Bulk API 2.0 job took the support-case text in a little over three minutes.9 Whether a chat agent's task needed that much access is not stated, and nothing is public on signing keys or a review against RFC 9700.
Limit which app permissions users can grant on their own.
No report says who granted Drift its Salesforce connection, or Drift Email its Google Workspace access. In Workspace, only accounts set up to integrate with Drift could be reached.11
5 more: 4 off the attack path, 1 not applicable
Keep authorization outside the model, and check every tool call against what the task allows.
The stolen tokens were used to run the exports,1 and Salesloft lists 20 addresses it calls confirmed malicious and three user-agent strings.10 On our reading those requests came from the intruder's own tools and never passed through the agent, so a check on the agent's tool calls would not have seen them.
- FENCE-CON-003
Not applicable
Require OAuth, tied to your identity provider, on every remote MCP server you run or approve. Tie each token to the person the agent acts for, or to a service with a named owner.
The reports describe no MCP server.
Avoid giving one agent identity the rights to both read private data and post where outsiders can see it. Where it needs both, block the post or put a person before it.
The chatbot ran on customers' websites,13 and the intruder used the integration's stolen credentials to read case data.2 Whether one identity held both rights is not stated, and the reports describe no use of a posting function.
Make each approval show the action and its details, and keep your own record of what was approved.
The reports describe no approvals.
Grant autonomy in levels, raise an agent's level only on evidence, and check again when its tools or setting change.
The reports give no autonomy level for the agent.
End
End it: the closest checks sit at the resource, after the fact
On our reading, the checks that meet this path most closely sit at the resource. After the breach Salesloft told partners that Drift traffic should come only from addresses Drift publishes;10 END-007 asks for that record before a breach. Cloudflare's dated timeline shows no alert before the August 23 notice.5 It then contained the integration and issued new secrets for every service by August 25.7
Test the off switch before you need it. Offboard a test agent, time how long each path into your systems takes to stop working, and list the tokens and sessions that survive.
Cloudflare used an off switch for real: by August 25 the Drift user was disabled, its client ID and secrets revoked, and Salesloft software removed.7 Nothing public says the switch had been tested beforehand, how long each path took to stop working, or whether any token survived.
Alert on credential behavior an agent was never meant to show. When a severe alert fires on an autonomous run, pause the run unless someone shows within a set time that the alert is false.
The intruder ran bulk queries and searched the data for keys and passwords.1,14 On our reading that is behaviour a chat integration was never meant to show. Cloudflare's dated timeline has no alert before the August 23 notice; its introduction says only, with no date, that it became aware of suspicious activity.5
At grant time, require a named business owner, a list of allowed actions, human approval for high-impact actions and an off switch.
No report describes what was required when the integration was granted: an owner, a list of allowed actions, approvals or an off switch.
Put agents in your access reviews, and re-check an agent's access when something changes, such as a new owner or a security event.
Cloudflare's dated response starts at the August 23 notice, not at Salesloft's August 20 revocation, which it had no sign concerned its environment.5 Whether the Drift integration was part of Cloudflare's access reviews is not stated.
Write agent logs from a component the agent cannot reach, and record both the agent and the person it acted for.
The intruder deleted its query jobs, yet GTIG says the logs were not impacted,1 and Cloudflare's log table times the intrusion to the second.9 Whether the integration could alter those logs, and whether entries named a person as well as the agent, is not stated.
Record where each agent's tokens are meant to be used from, such as its workload or network. When a valid token is used from anywhere else, alert an owner, have a control outside the agent revoke it within a set time, and check at the resource that it stops working within that time, while it would otherwise still be valid.
On our reading, this is the check that meets the path most closely. After the breach Salesloft told partners that Drift traffic should come only from addresses Drift publishes, and that an authenticated Drift-token connection from anywhere else is suspect.10 Whether Cloudflare kept such a record before the breach is not public; its dated timeline has no alert before the notice.5
When a vendor or integration on your inventory (FENCE-FIND-001, FENCE-FIND-003) is reported compromised, by its own notice or a provider's advisory, alert its owner, and within a set time have a control outside it revoke every token and grant it holds into your systems, and have the owners on its list (FENCE-FIND-006) revoke each credential that cannot be revoked in one place where it is held. Check at the resource that each stops working within that time, while it would otherwise still be valid.
Cloudflare's response did some of what this check asks. Notice reached it on August 23;5 containment and new secrets for every service followed by August 25.7 Salesloft revoked OAuth tokens centrally and left API keys to each provider.8 No set time and no check at the resource are reported.
Give each agent log entry identifiers that lead back to the token, grant and run behind the action, never the token itself.
GTIG points customers to Event Monitoring, the Drift connected app's authentication activity and the UniqueQuery events that record each query.1 Whether those entries lead back to a token and a grant is not stated.
3 more, off the attack path
When the person an agent acts for is flagged, end the agent's sessions too.
The reports do not say the integration acted for any person.
List in each agent's grant the credentials it is approved to hold. When a credential is added to the agent's identity outside the grant process, alert an owner, have a control outside the agent remove it and revoke the tokens it got within a set time, and check at the resource that they stop working within that time, while they would otherwise still be valid.
The reports describe no credential added to the integration's identity at Cloudflare; the intruder used tokens taken from Drift's AWS environment.4
When an agent or app on neither your agent register (FENCE-FIND-001) nor your integration list (FENCE-FIND-003) gets a token or a user's consent, alert an owner, have a control outside the app revoke its grant and tokens within a set time, and check at the resource that they stop working within that time, while they would otherwise still be valid.
The reported attack used an existing integration's tokens;1 no new app gaining a token or a user's consent is described.
Add-on · Payments
All 3 not applicable
- FENCE-TRX-001
Not applicable
Treat any agent that pays or releases funds as a privileged identity, with its own credential, a spending scope tied to the task, and a way to revoke it.
The reports describe no payments by the agent.
- FENCE-TRX-002
Not applicable
Keep the signing key and the running total outside the agent.
The reports describe no payments by the agent.
- FENCE-TRX-003
Not applicable
Know what each payment scheme proves. Telling an agent from a bot does not prove what the cardholder allowed.
The reports describe no payments by the agent.
What this edition does not cover
Secrets inside the data an agent may read. GTIG assesses, and Salesloft's initial findings say, that credentials were the intruder's main aim;1,14 Cloudflare found 104 Cloudflare API tokens in the stolen case text.3 No FENCE requirement covers secrets inside records an agent can reach.
The vendor's own estate. The Risky Business hosts describe the breach as the theft of tokens Salesloft held so its AI systems could work with customers' Salesforce data, Google Workspace, S3, Azure and OpenAI tokens alongside.6 FENCE meets this path only once a token is used; it cannot see how a vendor keeps its tokens.
A notice that does not say compromised. Salesloft's first notice said it had revoked the Drift-to-Salesforce connections and asked admins to re-authenticate;8 Cloudflare's dated response starts three days later.5 We would add a vendor's mass revocation as a trigger in a later edition; today END-010 starts at "compromised".
Scope that grows. Salesloft first said customers without the Salesforce integration were unaffected, then told all Drift customers to treat any Drift integration and its data as potentially compromised.13 GTIG's advice widened on August 28 to every token stored in or connected to Drift.11 END-010 reads as one event; re-running it on each widening is our recommendation, not yet the method.
One caution: FIND-003, FIND-006, END-007, END-010 and END-011 already cite this incident on the FENCE page; the case shows they meet the path, not that they are right.
›Evidence: 14 claims from 4 briefings
- 1BI-FIND-2025-0664 · GTIG traced Salesforce data theft to compromised Salesloft Drift tokens, then widened the scope to Workspace mailGoogle Threat Intelligence Group (GTIG) says that from as early as Aug 8, 2025 through at least Aug 18, an actor it tracks as UNC6395 used compromised OAuth tokens of the Salesloft Drift third-party app to export large volumes of data from numerous corporate Salesforce instances. Among the queries the post prints are row counts of the Account, Opportunity, User and Case objects, a pull of up to 20 active users' names, emails, phone numbers and last-login dates, and a pull of up to 10,000 Case records. GTIG assesses the main aim was credentials: the actor then searched the data for AWS access keys (AKIA), passwords and Snowflake-related tokens. It deleted its query jobs, but GTIG says the logs were not impacted, and it points Salesforce customers to Event Monitoring, the Drift Connected App's authentication activity and the UniqueQuery events that record executed queries. GTIG says the issue does not stem from a vulnerability in the core Salesforce platform; the post does not say how the tokens were obtained.
- 2BI-FIND-2025-0666 · A stolen Drift credential opened Cloudflare's support cases, which held 104 Cloudflare API tokensCloudflare says an actor it calls GRUB1 got into its Salesforce tenant at 22:14 UTC on August 12, 2025, using stolen OAuth credentials of the Salesloft Drift chat agent's Salesforce integration. What the actor reached was the text of Salesforce case objects: subject lines, the body of case correspondence and customer contact details, but no attachments or files. Cloudflare says it validated that limit through reviews of integrations, authentication activity, endpoint telemetry and network logs, and that no Cloudflare services or infrastructure were compromised. The post does not say what kind of token the integration used or what Salesforce access its Drift user account held.
- 3BI-FIND-2025-0670 · A stolen Drift credential opened Cloudflare's support cases, which held 104 Cloudflare API tokensCloudflare scanned the support-case text stolen from its Salesforce tenant in the Salesloft Drift breach with tools built on regex, entropy and pattern matching, found 104 Cloudflare API tokens, and rotated all of them; it says it has identified no suspicious activity on them. It looked because customers sometimes paste keys, logs or other sensitive data into case text, and because it believes the actor meant to harvest credentials and customer information for later attacks. The scan was built to find likely Cloudflare secrets, so we read the 104 as a count of Cloudflare's own tokens only. The post gives no count of other secrets and tells customers to treat anything shared in a case as compromised. Affected customers were notified on September 2 by email and dashboard banner.
- 4BI-FIND-2025-0671 · Salesloft says the Drift intruder was in its GitHub account months before the Salesforce data theftSalesloft's update as of September 6, 2025, relaying Mandiant's findings, starts the intrusion in Salesloft's own GitHub account. From March through June 2025 the actor was in that account, where it downloaded content from multiple repositories, added a guest user and set up workflows. It then accessed Drift's AWS environment, obtained OAuth tokens for Drift customers' technology integrations, and used them to reach data through those integrations. The September 30 summary puts the whole intrusion between March 22 and September 5, 2025, and adds API calls from Tor and anonymizing proxies, GitHub personal access tokens used for reconnaissance and secret enumeration, and the theft of environment-variable secrets and code repositories. Neither the September 6 update nor the September 30 summary says how the actor first got into GitHub, or how it got from GitHub to AWS.
- 5BI-FIND-2025-0668 · A stolen Drift credential opened Cloudflare's support cases, which held 104 Cloudflare API tokensSalesloft revoked Drift-to-Salesforce connections across its customer base on August 20 and posted a notice; Cloudflare says it had not yet been notified and had no sign the action concerned its environment. Salesforce and Salesloft notified Cloudflare on August 23, six days after the August 17 export job by our count. In the post's dated timeline, Cloudflare's response starts at that notice: first the vendors' recommended containment steps, then its own investigation from August 25. No earlier Cloudflare alert appears there; the introduction says only, without a date, that Cloudflare "became aware of suspicious activity".
- 6BI-FIND-2025-0299 · Salesloft's store of customers' OAuth tokens was stolen, and Risky Business sees few fixes customers controlDrawing partly on Cloudflare's own write-up, Patrick Gray and Adam Boileau describe the Salesloft Drift breach as the theft of OAuth bearer tokens that Salesloft held for its customers so its AI systems could work with their Salesforce data. The hosts add that the stolen set also held tokens for other services, Google Workspace, S3, Azure and OpenAI among them. The exposure went one step further: Cloudflare's support tickets, in systems integrated with Salesloft, held tokens to some of its customers' own Cloudflare services, which Cloudflare traced and rotated. Gray doubts that Cloudflare's planned weekly rotation of such tokens would have helped in this case, and he says that when they spoke, how the tokens left Salesloft was still unknown.
- 7BI-FIND-2025-0669 · A stolen Drift credential opened Cloudflare's support cases, which held 104 Cloudflare API tokensBy August 25 Cloudflare had disabled the Drift user account, revoked its client ID and secrets, purged all Salesloft software and browser extensions from its systems, and rotated credentials for the other third-party services connected to its Salesforce environment. It disconnected all third-party integrations from Salesforce, issued new secrets for all services and set up a process to rotate them weekly. Between August 26 and 29 it methodically re-onboarded the terminated integrations, each with new secrets and what it calls stricter security controls.
- 8BI-FIND-2025-0673 · Salesloft says the Drift intruder was in its GitHub account months before the Salesforce data theftSalesloft revoked access in more than one place, and split the work by credential type. The first incident notice says it revoked the connections between Drift and Salesforce and asked Drift admins to re-authenticate; the update giving the August 8 to 18 window says that, with Salesforce, it revoked all active access and refresh tokens for the Drift application. Salesloft recommended that customers who connect Drift to third-party applications with their own API keys revoke those keys and reconnect with new ones, while "OAuth applications are being handled directly by Salesloft." The notice announcing Drift's shutdown from Friday, September 5, 2025 at 6 a.m. ET says Salesloft rotated all centrally managed client keys, which "invalidated all affected tokens", and that API-key revocations will need to be done directly within each third-party provider's application.
- 9BI-FIND-2025-0667 · A stolen Drift credential opened Cloudflare's support cases, which held 104 Cloudflare API tokensCloudflare's log table times the intrusion to the second, from a failed token check at 11:51 UTC on August 9 to a Salesforce Bulk API 2.0 job that ran from 11:11:56 to 11:15:18 on August 17 and took the support-case text. GRUB1 deleted that job at 11:15:42, and Cloudflare saw no activity from it after that day. The post does not settle when theft began: its summary says data was exfiltrated between August 12 and 17, its August 13 narrative says the actor "stole customer data" from case objects from 19:33, and the table calls that day's first broad case query "reconnaissance via bulk record retrieval". It also describes the August 9 probe two ways. The prose says the actor tried a Cloudflare-issued API token "to the Salesforce API" with a request to client/v4/user/tokens/verify, while the table says the check ran against a Cloudflare customer tenant, and the post calls the token's source unclear.
- 10BI-FIND-2025-0675 · Salesloft says the Drift intruder was in its GitHub account months before the Salesforce data theftSalesloft's detection advice to integration partners turns on where a token is used from. Drift integration traffic should come only from a published set of source IPs that Drift owns and operates, and Salesloft says any "successfully authenticated connections utilizing Drift tokens" from another address should be considered suspect, investigated and reported to it. The partner FAQ also lists 20 IP addresses it calls confirmed malicious and three user-agent strings: python-requests/2.32.4, Salesforce-Multi-Org-Fetcher/1.0 and Python/3.11 aiohttp/3.12.15.
- 11BI-FIND-2025-0665 · GTIG traced Salesforce data theft to compromised Salesloft Drift tokens, then widened the scope to Workspace mailOn Aug 20, Salesloft, working with Salesforce, revoked all active access and refresh tokens with the Drift app, and Salesforce removed Drift from its AppExchange. GTIG's update of Aug 28 then said the compromise was not limited to the Salesforce integration. That day its investigation confirmed the actor had also compromised OAuth tokens for the "Drift Email" integration, and on Aug 9 had used them to access email from "a very small number of Google Workspace accounts"; only accounts set up to integrate with Salesloft Drift could have been reached. Google revoked the tokens granted to the Drift Email app and disabled the integration between Workspace and Salesloft Drift, and says Workspace itself was not compromised. GTIG now advises all Drift customers to treat every authentication token stored in or connected to the Drift platform as potentially compromised. The body of the post still carries Salesloft's earlier word, based on the data it had at the time, that customers who do not integrate with Salesforce were not impacted.
- 12BI-FIND-2025-0677 · Salesloft says the Drift intruder was in its GitHub account months before the Salesforce data theftMandiant's September 30 summary lists hardening across the Drift application, its AWS infrastructure and GitHub, from code vulnerability analysis to logging and CI/CD workflow review. The items that touch credentials and access include credential rotation in the Drift application, an IAM role and policy review, secrets-prevention tooling with rotation of secrets found in code, and eliminating GitHub personal access tokens and external collaborators as means of access. The update announcing Drift's return on Tuesday, September 16 says Salesloft rotated customer and internal credentials, including OAuth tokens, employee passwords and other application secrets, improved its logging, upgraded authentication for privileged users, shortened session lengths for internal systems, and limited privileges to the lowest levels needed.
- 13BI-FIND-2025-0674 · Salesloft says the Drift intruder was in its GitHub account months before the Salesforce data theftSalesloft's account of who was exposed grew wider. The undated update that gives the August 8 to 18 window, one of the oldest entries on the page, says the incident did not affect customers who do not use the Drift-Salesforce integration. Other undated entries report that Salesforce disabled the Drift integration with Salesforce, Slack and Pardot, and that the Salesforce-Salesloft integration was paused, a precaution Salesforce initiated; Salesloft says it saw no evidence of malicious activity on its own platform. Salesloft also took Drift itself offline from Friday, September 5, 2025 at 6 a.m. ET until Tuesday, September 16. While it was down, the Drift chatbot was off customers' websites, Drift Fastlane and Drift Email were not accessible, and the JavaScript snippet on customers' sites would not load Drift. An entry posted after the September 6 update tells all Drift customers to treat "any and all Drift integrations and related data as potentially compromised."
- 14BI-FIND-2025-0672 · Salesloft says the Drift intruder was in its GitHub account months before the Salesforce data theftIn an undated update, Salesloft says that from August 8 to August 18, 2025 a threat actor used OAuth credentials to exfiltrate data from its customers' Salesforce instances, and that initial findings showed the actor's primary objective was to steal credentials, with a focus on sensitive information such as AWS access keys, passwords and Snowflake-related access tokens. The actor queried Salesforce objects including Cases, Accounts, Users and Opportunities. Of the two sample queries Salesloft prints, one pulls the description, subject and comments of Case records and the other mines a Case field for known secret patterns.
A worked case reads one incident with hindsight, from public reporting only. It shows how FENCE applies, not what any party knew at the time.
Read the FENCE method Assess one agent against the same requirements