Agent incident database
ForcedLeak: a web form entry made Salesforce Agentforce leak CRM data
25 Sept 2025Salesforce · Agentforcedata exfiltrationdisclosed vulnerability
Noma Security showed that instructions typed into a public web-to-lead form could later make Salesforce's Agentforce send CRM data to an attacker. The route out used a domain Salesforce still trusted but had let expire.
How it unfolded
Each step on its own looked like ordinary agent work. The harm came from the sequence.
- 1
Stored a lead submitted through a public web form, with hidden instructions
- 2
Read the lead record when an employee asked about it
- 3
Model followed the instructions hidden in the lead
- 4
Queried other CRM records, such as lead email addresses
- 5
Output an image tag that sent the data to an attacker-owned, still-whitelisted domain
harm
Which policy would have stopped this?
In our reading of the reports, these policies from the open library stop the harmful step before it runs.
- Restrict outbound to approved domains: The data left in a request to a host nobody approved. An outbound allowlist blocks that step.
- Block external send after sensitive data read: The request carried CRM data the agent had just read.
Open Restrict outbound to approved domains in the playground and build the path yourself.
This is our reading of public reporting. We haven't run it against the original systems.
What happened
Noma Labs reported the chain (CVSS 9.4) to Salesforce on 28 July 2025 and disclosed it on 25 September 2025. An attacker submits a lead through Salesforce's Web-to-Lead form, placing instructions in the 42,000-character Description field, where they are stored as ordinary CRM data. When an employee later asks Agentforce about that lead, for example to draft a follow-up email, the agent processes the hidden instructions and queries other CRM data. It then outputs an HTML image tag pointing to an attacker URL with the data in a query parameter. The content security policy still allowed a domain, my-salesforce-cms.com, that had expired; the researchers bought it for about $5 so the request would pass.
Outcome
Salesforce re-secured the expired domain and on 8 September 2025 enforced Trusted URLs for Agentforce and Einstein AI so agent output cannot be sent to untrusted URLs. Salesforce told The Register it had released patches but declined to say whether the flaw had been abused.
Sources
- ForcedLeak: AI agent risks exposed in Salesforce Agentforce, Noma Security, 2025-09-25
- Salesforce Agentforce tricked into leaking sales leads, The Register, 2025-09-26
- Salesforce Patches Critical ForcedLeak Bug Exposing CRM Data via AI Prompt Injection, The Hacker News, 2025-09-25
A summary of public reporting. The step-by-step timeline is our reading of those sources; where they differ, we say so above.
Your agents
Could this happen to yours?
For developers
Build a path yourself
Recreate a sequence in the playground and see how policies treat it.