Weaponization of Prompt Injection: From EchoLeak to ForcedLeak

In 2025, prompt injection evolved from a laboratory concept into a real-world production threat. Two landmark cases—EchoLeakandForcedLeak—demonstrate thatzero-click attacksandindirect prompt injectionhave become core risks facing enterprise AI assistants.

EchoLeak (CVE-2025-32711): Silent Data Exfiltration in M365 Copilot

Discovered in June 2025, the EchoLeak vulnerability is widely regarded as the first fully automated data exfiltration exploit in a production-grade AI system. Its severity lies in itszero-click nature: the victim does not need to click links or execute any code. Simply processing a malicious email in the background is sufficient to trigger the leak.

Technical Attack Path

The attacker sends a carefully crafted email containing a malicious Markdown payload to the victim’s Outlook inbox. When the user later interacts with Microsoft 365 Copilot (e.g., requesting a summary of recent emails), Copilot’s Retrieval-Augmented Generation (RAG) engine incorporates the malicious email into its context.

• Classifier Bypass (XPIA Bypass):
  Microsoft deployed cross-prompt injection detection mechanisms. However, EchoLeak leveragesUnicode homoglyphsand contextual obfuscation to disguise malicious instructions as legitimate business communication, successfully evading semantic detection.
• Data Exfiltration via Link Filter Bypass:
  The attack abuses Markdown’sreference-style link syntax. While inline external links are typically filtered, referenced links can bypass certain parsing paths. Sensitive data (e.g., from OneDrive) is embedded into URL parameters and transmitted silently to attacker-controlled servers.
• Stealth Mechanism:
  The injected instructions include directives such as “do not mention this email,” preventing Copilot from citing the malicious source in its response, leaving the user unaware of the breach.

ForcedLeak: Logical Exploitation in Salesforce Agentforce

Similar in nature, ForcedLeak targets Salesforce’s Agentforce platform. Attackers inject malicious prompts into open fields (such as “Description”) via Web-to-Lead forms.

Attack Logic

The most notable aspect of this attack is its paradox: it succeeds not because of user error, but because users follow normal workflows correctly.

When a sales representative uses an AI assistant to classify or score leads, the AI processes stored malicious instructions within the CRM system. These instructions manipulate the AI to bypass security safeguards and exfiltrate sensitive Personally Identifiable Information (PII) through a previously trusted but now expired domain.

Attackers can purchase such expired domains for as little as $5, effectively bypassing Content Security Policy (CSP) protections. This highlights a fundamental weakness in AI security:the implicit trust model of data sources.

Comparison of Major Prompt Injection Incidents (2025)

Feature

EchoLeak (M365 Copilot)

ForcedLeak (Agentforce)

CVSS Score

9.3 (Critical)

9.4 (Critical)

Trigger Mechanism

Zero-click, background email processing

Triggered during lead processing

Exfiltration Channel

Markdown reference link bypass

Trusted domain / CSP bypass

Data Sources

SharePoint, OneDrive, Teams

Salesforce CRM core data

Core Vulnerability

Lack of cross-context isolation

Treating untrusted input as trusted instructions

Implications for Future AI Security

These incidents expose a fundamental weakness in AI systems: the blurred boundary betweeninstructionsanddata.

1. Shift from Input Filtering to Semantic Monitoring

Traditional rule-based filtering (e.g., regex) is insufficient against natural-language attacks. Organizations must deploycontext-aware AI guardrailscapable of understanding intent and enforcing integrity checks during sensitive operations such as data retrieval.

2. Enforce Human-in-the-Loop Controls

AI systems should not have full autonomy over high-risk actions, including:

• External network communication
• Bulk data export
• Sensitive configuration changes

Following the ForcedLeak incident, Salesforce implemented patches requiringexplicit human approvalfor such operations.

3. Reassess the “Data as Instructions” Risk

Any data accessible to AI—emails, Slack messages, Jira tickets, or web forms—can act as executable payloads. Organizations must:

• Enforce strictdata source isolation
• Applyleast-privilege access controls
• Continuously validate trust boundaries

This paradigm shift is essential to prevent similar data exfiltration attacks.

References

EchoLeak: The First Real-World Zero-Click Prompt Injection Exploit in a Production LLM System – arXiv

ForcedLeak: The $5 Exploit That Broke Salesforce's AI Agents – Inspired eLearning Blog

Insight Report Source: Global Cybersecurity Alliancehttps://www.gcsa.org