R-AC-08 Agent Communication & Interoperability DAMAGE 3.4 / High

Human Channel Impersonation

Agents communicating through human channels (Slack, Teams, email) can be indistinguishable from human participants. No institutional policy requires agents to identify themselves.

The Risk

When agents communicate through the same channels as humans (email, Slack, Teams, SMS), users cannot reliably distinguish agent outputs from human messages. An agent responding to a question is indistinguishable from a human responder. This creates impersonation and trust vulnerabilities.

A user receives a message in Slack claiming to be from "Project Manager Chris" providing project status. The user assumes the message is from Chris (a human colleague) and trusts the information. But the message came from an agent trained on Chris's writing style. Additionally, humans may behave differently when interacting with agents vs. humans, and the lack of transparency about agent vs. human enables manipulation.

How It Materializes

A financial services company deploys an agent-based customer support system. When customers email support@company.com, the email goes to an agent that responds to routine questions (account balance, transaction history, payment processing). The agent's responses are formatted like human support emails, with signature "Customer Support Team."

A customer emails asking about a late charge on their account. Agent-Support reads the email and generates a response explaining the charge and offering a refund. The response is professional, empathetic, and accurate. The customer reads the response and replies, continuing the conversation.

However, the customer does not know the conversation is with an agent. The customer assumes they are communicating with human support. Over 5 emails, the customer shares sensitive information (account numbers, prior payment difficulties) because they believe they are communicating confidentially with a human support representative. The agent's responses are logged to a system database (unlike human support emails which might be deleted). The sensitive information is now in a database accessible to many employees.

Additionally, an adversary could exploit this impersonation. If the adversary can inject messages into the support channel, customers will trust messages from "Customer Support Team" without realizing the message is from an attacker.

DAMAGE Score Breakdown

Dimension Score Rationale
D - Detectability 2 Agent messages are often indistinguishable from human messages. Users may not realize they are communicating with an agent.
A - Autonomy Sensitivity 2 Affects all agent communication types. Transparency and labeling would address risk.
M - Multiplicative Potential 2 Affects user interactions with agents. Impersonation risk scales with number of agent interactions.
A - Attack Surface 3 Human communication channels are attack surfaces. Compromised channels can be used to impersonate agents.
G - Governance Gap 3 Institutions may not have policies requiring transparency about agent vs. human communication.
E - Enterprise Impact 2 Affects user trust and information security. Does not directly impact financial transactions unless users make decisions based on impersonation.
Composite DAMAGE Score 3.4 High. Requires dedicated controls and monitoring. Should not be accepted without mitigation.

Agent Impact Profile

How severity changes across the agent architecture spectrum.

Agent Type Impact How This Risk Manifests
Digital Assistant Low Digital assistants are typically labeled as assistants in UI. Users know they are interacting with AI.
Digital Apprentice Low-Med Agents may be labeled in communication, but labeling might not be obvious.
Autonomous Agent Medium If agent communicates in human channels without labeling, impersonation risk is high.
Delegating Agent Medium If delegating agent invokes tools that communicate in human channels, impersonation can occur.
Agent Crew / Pipeline Medium Multiple agents in crew may communicate in human channels, creating impersonation risk if not labeled.
Agent Mesh / Swarm High Mesh agents may communicate in human channels directly. Impersonation risk is high.

Regulatory Framework Mapping

Framework Coverage Citation What It Addresses What It Misses
FTC AI Guidance Partial Deceptive AI Practices Prohibition on deceptive AI use. Requirement for transparency about agent vs. human.
GDPR Article 21 Partial Right to Object Transparency about automated decision-making. Requirement to disclose when communication is from agent.
SEC Disclosure Rules Partial Plain English Requirements Clear communication to investors. Disclosure of agent-generated content.
NIST CSF Minimal Cybersecurity governance. Human channel security and agent impersonation.

Why This Matters in Regulated Industries

In financial services, customers rely on clear communication about who they are communicating with. If a customer believes they are communicating with human support but are actually communicating with an agent, they may share sensitive information or make decisions based on incorrect assumptions about the conversation's confidentiality.

Additionally, regulatory frameworks increasingly require transparency about AI involvement in customer interactions. The FTC has issued guidance against deceptive AI practices, including impersonation.

Controls & Mitigations

Design-Time Controls

  • Require transparent labeling of agent communications. Every agent message in human channels must clearly identify itself as agent-generated.
  • Implement distinct naming or email addresses for agent-generated communications. Instead of "support@company.com," use "bot-support@company.com."
  • Use Component 7 (Composable Reasoning) to enable agents to reason about transparency requirements and include transparency labels in all outputs.
  • Design agents to escalate to human when sensitive information is shared. If customer shares account numbers, agent should escalate and notify customer.

Runtime Controls

  • Log all agent communications in human channels. Maintain audit trail of what agents said and to whom.
  • Implement monitoring for agent impersonation. Flag messages claiming to be from known humans but actually generated by agents.
  • Monitor human channels for anomalous activity. If agent accounts are hijacked or if fake agents are injected into channels, detection systems should flag unusual patterns.

Detection & Response

  • Conduct user surveys to assess awareness of agent vs. human communication. If significant percentage of users believe they communicate with humans when actually communicating with agents, increase transparency.
  • Analyze complaints about agent communications. If users report feeling deceived or misled, investigate whether transparency was insufficient.
  • Implement incident response for impersonation incidents. If attackers successfully impersonate agents, investigate and remediate.

Related Risks

Address This Risk in Your Institution

Human Channel Impersonation requires architectural controls that go beyond what existing frameworks provide. Our advisory engagements are purpose-built for banks, insurers, and financial institutions subject to prudential oversight.

Schedule a Briefing