New Attack Chain Uses AI Agents to Slip Phishing into Slack

Researchers have identified a technique dubbed Salesbleed that leverages agentic AI to inject malicious instructions from external web content into trusted internal messaging platforms. The attack can traverse multiple applications, ultimately delivering phishing messages through Slack without triggering typical security alerts. The method highlights risks posed by AI agents that process untrusted data.

The Salesbleed technique exploits a growing weakness in modern workplaces: AI agents that automatically process and act on external data. By embedding malicious instructions within web content, attackers can hijack these agents, turning them into unwitting conduits for phishing. Because the agents operate within trusted platforms like Slack, the fraudulent messages appear legitimate and bypass conventional email-focused security filters.

The attack's multi-application traversal is particularly concerning, as it demonstrates how a single compromised data stream can propagate across an organization's entire digital ecosystem. This research underscores a fundamental tension in enterprise AI adoption—the same automation that boosts productivity also expands the attack surface. Security teams must now consider whether their AI tools can be trusted to handle untrusted input safely.