Humen — agentic threat model

Not certain from the listing — likely relies on commercial LLMs for generating outbound copy. Primary threats include prompt injection leading to brand damage or the generation of highly convincing, unauthorized social engineering templates.

Not certain from the listing — processes prospect activity signals and ideal customer profiles. Threats include CRM data exfiltration, poisoning of lead databases to redirect outreach, and unauthorized access to sensitive customer interaction histories.

The agent orchestrates multi-step workflows (lead generation, copy creation, outreach, and qualification). A critical threat is indirect prompt injection, where a prospect's malicious reply manipulates the agent's internal state or triggers unauthorized tool actions.

Not certain from the listing — deployed as a closed-source SaaS. Main threats involve the secure storage of API keys and OAuth tokens for third-party platforms (CRMs, email servers, LinkedIn) which, if compromised, grant attackers direct access to communication channels.

Not certain from the listing — requires robust guardrails to monitor outbound copy before transmission. Without strict observability, the agent could autonomously send non-compliant, offensive, or highly repetitive spam without detection.

Fully autonomous outbound outreach carries severe compliance risks under GDPR, CAN-SPAM, and platform-specific terms of service (e.g., automated messaging bans). The lack of explicit human-in-the-loop controls increases regulatory exposure.

Not certain from the listing — currently operates as a standalone SDR digital worker. However, future integrations with other automated 'Humen' digital workers could introduce cascading trust failures and multi-agent coordination vulnerabilities.