AI Threat Watch

Case copilot-agentforce-injectionPrompt injectionWorldwide

Prompt injection leaks data from Microsoft Copilot Studio and Salesforce Agentforce agents

Researchers demonstrated indirect prompt injection that exfiltrated data from enterprise agents built on Copilot Studio (CVE-2026-21520) and Agentforce despite earlier patches.

HighAttack on AIStatus: PatchedVerified case file

Open in the interactive mapCite this case

Reported
Apr 15, 2026
Location
Worldwide
Target
Copilot Studio, Salesforce Agentforce
Severity
High
Triggered by
Security researchers at Capsule Security (responsible disclosure)
Actor type
Security researchers
Reported loss
None reported
CVE
CVE-2026-21520

What happened

Researchers demonstrated indirect prompt injection that exfiltrated data from enterprise agents built on Copilot Studio (CVE-2026-21520) and Agentforce despite earlier patches.

Impact: Demonstrated exfiltration of SharePoint list data by email and misuse of Agentforce workflows; Microsoft safety filters flagged but did not stop the payload; no confirmed real-world victims

Who triggered it

Actor: Security researchers at Capsule Security (responsible disclosure)

Attribution: Self-attributed research by Capsule Security; Microsoft assigned CVE-2026-21520 (ShareLeak); Salesforce confirmed addressing PipeLeak

Motive: Security research and responsible disclosure

How it happened

Attack vector: Indirect prompt injection via ordinary business inputs: a SharePoint form comment field (Copilot Studio) and a public lead form (Agentforce) carrying fake system instructions

Root cause: Untrusted form input was concatenated into agent system prompts with no separation between trusted instructions and user data, so agents followed injected commands through authorized tools

Timeline

  1. Microsoft patches the Copilot Studio flaw server-side
  2. CVE-2026-21520 published (CVSS 7.5)
  3. Capsule Security publicly discloses ShareLeak and PipeLeak alongside its launch

Response

Microsoft mitigated internally with no customer action required; Salesforce said it addressed the issue; researchers warned the attack class remains open

Current status: Patched

Sources (open in a new tab)

Cite this case

Prompt injection leaks data from Microsoft Copilot Studio and Salesforce Agentforce agents. AI Threat Watch, case copilot-agentforce-injection, Apr 15, 2026. https://aithreat.io/cases/copilot-agentforce-injection/

Permanent link: https://aithreat.io/cases/copilot-agentforce-injection/

This case file in other languages: EnglishFrançaisEspañolالعربية中文