# Public forms as an attack path: ShareLeak and PipeLeak

> A security firm showed that text typed into an ordinary web form could steer a Copilot Studio agent and a Salesforce Agentforce agent into emailing customer data to an outsider. The cases are from April 2026 and fit the lethal trifecta.

Source: https://conversationalaidesign.com/news/form-based-prompt-injection-copilot-studio-agentforce.html. Author: Nelson Abacherli Kawakami. Published November 17, 2026. Reported.

## What was reported

Capsule Security published two findings on 15 April 2026.

- **ShareLeak (Microsoft Copilot Studio).** A crafted payload in a SharePoint form field, such as comments, was joined to the agent's instructions and overrode them. The agent pulled customer records from connected SharePoint lists and sent them by email. Microsoft assigned **CVE-2026-21520** (CVSS 7.5) and patched it in January 2026. Reports say its safety checks flagged the request, yet the data still left.
- **PipeLeak (Salesforce Agentforce).** Instructions hidden in a public lead form ran when an employee asked the agent to process the lead, and the agent emailed CRM data out. Salesforce said it remediated the specific scenario and turned on human confirmation by default for email actions. Capsule said the email path was still open on custom topics. **That dispute is not settled in what we read.**

## Why it matters for support design

Both cases are the **lethal trifecta** in a plain form: private data, untrusted input and a way to send data out. Nothing here needed a clever model attack. A form field was enough.

## What to check in your own design

- List every **public input** that your agents read: forms, tickets, emails, chat. Treat each as untrusted.
- Limit **outbound actions** such as email to approved domains, and ask a person to confirm sending data to anyone else.
- Give each agent **least-privilege access** to the data it needs for one job.
- Do not rely on the model flagging suspicious text. Put the block in the tool, not in the prompt.

## How this was verified

On 17 November 2026 two trade articles were read, and they agree on the facts above. The researchers' own reports and the vendors' advisories were not read, so this is **reported**, not verified. Earlier roundups placed these cases in October; the original disclosure was in April.

## Sources
- CSO: Copilot and Agentforce fall to form-based prompt injection tricks (15 Apr 2026): https://www.csoonline.com/article/4159079/copilot-and-agentforce-fall-to-form-based-prompt-injection-tricks.html
- VentureBeat: Microsoft and Salesforce agents, prompt injection, remediation playbook: https://venturebeat.com/security/microsoft-salesforce-copilot-agentforce-prompt-injection-cve-agent-remediation-playbook
