how it happened
- whoanyone on the internetno account, no login
- the targeta lead record with hidden instructionssits in the CRM like any other lead
- thenthe agent obeys the lead textpulls CRM data into an image link
- the resultthe trusted domain had expiredresearchers bought it and caught the data
- how it stoppedfixed before disclosureSalesforce enforced trusted URLs from 8 September 2025
what everyone assumed
A contact or lead form is the most harmless thing on a company website. It collects a name, an email address and a message, and somebody in sales reads it. The worst a stranger could do was send spam or a rude note, because a person decided what happened next. Plenty of careful companies treat what comes through that form as data to be read, and nothing more.
That held while only people read the leads. It stopped holding once an AI agent read them too, because the agent takes in the text of the lead together with the employee’s request, and it can act on both.
what happened
This was a researchers’ demonstration. Noma Labs, the research team of the security vendor Noma Security, built the attack in a setup it configured for the purpose (“For this research, we enabled Salesforce’s Web-to-Lead feature”, in Noma’s words). No source opened for this page reports any real company’s data being taken, and none reports the method being used by attackers.
The mechanism, as Noma’s write-up of 25 September 2025 describes it:
- The door anyone can use. Web-to-Lead is a public form that creates lead records in Salesforce. Noma used the lead’s Description field, which it says holds 42,000 characters, to hide instructions inside an otherwise normal-looking submission.
- The routine question. Later, an employee asks the Agentforce agent to look at that lead and answer the person’s questions. Noma lists predictable employee questions among the causes: the attacker can count on someone asking.
- The agent obeys the lead. The agent read the hidden text as instructions. In Noma’s payload those instructions asked for CRM data, including lead email addresses, and told the agent to put it into the address of an image.
- The trusted list with a hole. Salesforce restricted which outside domains such content could load from. One domain on that list had expired. Noma says it bought the domain for $5, so the image request, carrying the data, arrived at a server Noma controlled.
Noma rated the chain at CVSS 9.4, a score it assigned itself; CSO Online reported that no CVE number had been issued yet at disclosure. Noma names four weak points: little limit on what the agent could query, no cleaning of fields the public can write to, an allowlist with an expired entry, and predictable employee questions.
step by step
- a public form feeds the CRMWeb-to-Lead turns whatever a visitor types into a lead record, and an AI agent can read those records for staff.
- July 2025researchers write instructions into the description fieldNoma Labs, in its own test setup, hid commands in a field that holds 42,000 characters, and the agent followed them when an employee asked about the lead.the assumption breaks here
- 28 July 2025Noma reports it to SalesforceThe data in the demonstration went to a domain on the agent's allowlist that had expired and that Noma had bought for $5.
- 8 to 25 September 2025Salesforce enforces trusted URLs, Noma goes publicEnforcement started on 8 September and the write-up followed on 25 September; no source reports real data taken.
how it surfaced
Noma reported the issue on 28 July 2025, and Salesforce acknowledged it on 31 July, according to Noma’s timeline. Salesforce’s own Help article says that from 8 September 2025 it began enforcing trusted URL allowlists for Agentforce and Einstein generative AI: a link to a domain that is not on the list or in the agent’s instructions is blocked. Noma says Salesforce also took back the expired domain. The article does not mention ForcedLeak or Noma; the connection between the report and the fix comes from Noma and CSO Online. Salesforce told CSO Online that prompt injection “remains a complex and evolving area”. Noma published on 25 September 2025.
The fix closed the image route. It did not change the fact that the public can write text an agent reads. On 15 April 2026 VentureBeat reported research by Capsule Security, another vendor, on a variant it calls PipeLeak: a payload in a public lead form steered an Agentforce agent into sending CRM data by email through the agent’s own permitted actions, with no volume limit seen and no notice to the employee. Salesforce told VentureBeat it had remediated the specific scenario and that human confirmation is on by default for agent email actions; Capsule said its retest showed the email route still worked on Custom Topics. That disagreement is open, and it is again research, with no reported victims.
In February 2026 Salesforce updated the same Help article: from 28 February 2026 the broad *.salesforce.com wildcard leaves the default allowlist, and admins are asked to audit their trusted URLs and list specific domains.
The same pattern shows up elsewhere in this record through other doors: an email read by Copilot and a tool’s description. What sets this one apart is who could use it: anyone who can reach your website, with no account at all. And the exit was a domain someone forgot to renew. Who controls your own domains is the subject of the guide on who can move your domain.
Once an AI agent reads your web form, a stranger’s message is also a set of instructions to something that can reach your CRM.
The guide on the agent’s permissions covers how to shrink what that agent can reach, and three new doors shows the same problem at other layers.
picture it at a company like yours
The company here is made up. The other two paths are not: they happened, and each links to the story.
- in this storythe day it mattersA stranger fills in the form with instructions written for the assistant, and a salesperson asks the assistant about that lead.what it costs youCustomer contacts leave through a link nobody looks at, and the first sign is someone else's write-up.
- or it goes like thisit happenedEchoLeak / Microsoft 365 Copilot, 2025howAn email with hidden instructions reaches an assistant that reads every inbox.what leavesanything the assistant can seewho ends up with itthe stranger who sent the email
- or it goes like thisit happenedMCP tool poisoning, 2025howA tool description tells the agent to send files somewhere else.what leavesfiles the agent can openwho ends up with itwhoever wrote the tool
sources
- Noma Security: ForcedLeak: AI agent risks exposed in Salesforce Agentforce, 25 September 2025 (vendor research)
- Salesforce Help: Update & Action Required: Changes to Default Allowlist for Agentforce (original article 8 September 2025, updated 13 February 2026)
- CSO Online: Vulnerability in Salesforce AI could be tricked into leaking CRM data, 25 September 2025
- VentureBeat: Microsoft patched a Copilot Studio prompt injection. The data exfiltrated anyway, 15 April 2026, updated 16 April 2026
your turn
what you can do about it
this week
List every AI feature that reads your public form submissions, and remove expired or broad entries from its allowed web addresses.
Ask whoever runs the CRM, helpdesk or chat tool which AI agents can read leads, tickets or form messages that strangers write. For each one, open the list of outside web addresses it may link to or load (in Salesforce, Setup, then Trusted URLs), delete any domain you do not recognise or no longer own, and replace wildcards with named domains. Then check that any agent action that sends email or calls an outside address asks a person first. If a stranger filled in your form tonight, what could your agent send out tomorrow?
go deeper
if you're not sure where to start
Run the offboarding checklist on one leaver. It covers every account and token to close when someone leaves.
open the checklist →
