I Investigated Contact Center Fraud (Here Is What I Found)

AI safety vs. human systems

The video argues that AI is not inherently safer than human systems—and in some cases, more capable models may be easier to socially engineer.

• Stanford research is referenced: smarter AI models with more “reasoning” can become more persuadable, because they try to complete the task and may “talk themselves into” accepting scams.
• In contact-center environments, attackers can use the same techniques that trick distracted employees to fool AI agents—which may then act on the organization’s behalf.

Insider risk in contact centers and BPO

Todd Schiller (cybersecurity expert and CEO/co-founder of PixieBrix) frames the broader issue as insider risk in:

• Contact centers
• Business process outsourcing (BPO)

He breaks insider problems into two categories:

• Insider mistakes, such as:
  • fat-fingering
  • wrong refunds
• Insider fraud, such as:
  • refund/price fraud
  • address/location fraud
  • loyalty/gift card fraud
  • exfiltration / snooping

Why contact-center workflows make it easier

Insider abuse often succeeds by exploiting the structure of contact-center workflows. Agents may need access to multiple internal systems (e.g., ERP/order management, CRM, logistics portals, and more).

Individually, each access permission may appear legitimate, but combining access across systems enables “business logic abuse”—using valid tools in an illegitimate sequence.

Outsourcing and growth-stage dynamics

Schiller notes two common situations that increase exposure:

• High-growth firms may lack a CISO and governance early, adding controls only later (e.g., around IPO or later funding).
• Legacy firms may rely on ERP/order systems not designed for fine-grained access controls, leading to workarounds.

Address/location fraud and detection challenges

He gives examples of address/location fraud that can be difficult to detect, including:

• exploiting geography and address patterns
• using knowledge of delivery routes

One example describes insiders timing thefts along UPS/FedEx routes—turning customer service into “porch pirate” behavior.

Detection is complicated by compliance constraints, including PCI DSS and the sensitivity of bill-to/ship-to addresses, which can create false positives, especially in dense cities.

How AI changes fraud and exfiltration threats

A major portion of the discussion ties AI to:

• fraud facilitation
• insider/exfiltration threats

Schiller argues that AI can make outsider attacks easier, such as:

• voice phishing
• deepfakes
• doctored images supporting fraudulent claims/refunds
• prompt-injection and social engineering of customer chatbots

Liability and automation boundaries

He also highlights legal/regulatory dynamics: if chatbots provide binding quotes in some jurisdictions, companies may face liability.

From a defense standpoint, he recommends:

• deciding carefully which flows are automated by AI
• using chatbots for safer tasks (e.g., password reset deflection)
• requiring stronger identity verification for higher-risk actions

PixieBrix: what it does and why

The video’s central “solution” claim is that PixieBrix protects AI agents and contact-center workflows by enforcing controls across systems that standard tools often fail to correlate.

PixieBrix is described as having three main components:

1. Control plane / policies Fine-grained policy controls by function, line of business, and seniority—addressing the problem that many systems effectively “give the keys to the kingdom.”

2. Browser extension Understands agent actions across browser tabs and maps them to activity across multiple applications (e.g., dialer, order management/ERP, logistics portals, reference tools).

3. Interaction/transaction matching + enforcement Determines whether an action is justified (e.g., whether an agent is handling the relevant customer ticket/record) and applies response levels from:

   • monitoring
   • warning
   • soft blocking (requesting attestation)
   • hard blocking

Enforcement example: gift card “coaching” fraud

Schiller illustrates how enforcement can stop insider “coaching” fraud in gift card workflows. In the scenario:

• an insider chats/calls until a new agent is connected
• guides the agent through an internal tool to execute a special gift card/refund flow
• monetizes it (e.g., converting into fungible high-value items)

PixieBrix can remove UI options and block URLs for certain groups, preventing the workflow entirely.

Browser-centric defenses against “dark patterns”

The video also highlights PixieBrix’s protections against deceptive UX aimed at AI/agents (“dark patterns”). Schiller references an open-source project—Agent Browser Shield—designed to apply multiple policies that hide or neutralize manipulative elements, such as:

• countdown timers
• inventory claims
• pre-checked add-ons

The goal is to prevent agents from being socially engineered into accepting options they shouldn’t.

Key recommendations emphasized

• Least privilege for agents Don’t give AI agents broad access. Restrict information, send destinations, and actions (he references a “lethal trifecta” concept).

• Redaction and DLP-style controls for generative AI When employees use chat AI tools, implement fine-grained controls such as:

  • blocking certain tools
  • redacting sensitive data (e.g., emails/phone numbers)
  • preventing risky prompts

• Correlate activity across systems Monitor not just single-app behavior, but cross-system sequences that enable fraud or data misuse.

Where this is being applied / who gains most

Schiller states the biggest benefits are in:

• E-commerce (high-value items; fast fraud impact)
• Telecom (snooping, SIM swapping)
• Travel & leisure (complex booking-change processes where mistakes can trigger large BPO penalties and delayed discovery)

Closing outlook

He mentions upcoming conference appearances, including:

• Black Hat (early August)
• a Boston cybersecurity conference (mid-August)
• major vendor conferences in Oct–Sept (e.g., Dreamforce/HubSpot)
• Contact Center Week Europe (early October, Amsterdam)

He frames the work as continuing to evolve with AI and contact-center adoption.

Presenters / contributors

• David Dean Marrow — host
• Todd Schiller — co-founder & CEO of PixieBrix; cybersecurity expert; PhD in Computer Science (University of Washington) and master’s in machine learning