AI-Enabled Fraud and What It Means for Audit
ISACA Tucson · November 19, 2026

Ryan Straight, Ph.D.

University of Arizona

Background

Ryan Straight in Edinburgh

  • Assistant Professor, Cyber Operations, College of Information Science
  • Lead, Human-AI Collaboration and Pedagogy, Arizona Cybersecurity Academy
  • Research: how people actually work with AI, and how attackers design what targets see

The Arup Deepfake Call

HK$200 million

15 transfers after one deepfake video call

“None of our internal systems were compromised.” (Arup, to CNN)

CNN (Magramo), May 16, 2024; Hong Kong police via RTHK.

Scope of the Talk

Auditing Your Own AI

ISACA's AAIA credential

The Attacker's AI

Aimed at staff, payments and sign-offs

ISACA, AAIA exam content outline (checked November 2026).

AI-Written Spear Phishing

Heiding et al., Expert Systems with Applications, 2026. 101 university volunteers, about 25 per group.

Reported Losses

$893 million
in FBI complaints that mentioned AI, 2025.
Probably an undercount.

FBI IC3, 2025 Internet Crime Report, pp. 39–40, 61.

Appraising Stolen Data

“a forensic intelligence analyst drafting a target assessment”

The VOLTA infostealer’s instruction to a commercial AI model, which ranks what each victim is worth

Cisco Talos (Fetterman), October 31, 2026. One sample analyzed.

ClickFix

  • The victim copies and runs the attack command
  • Browser download protections never fire
  • The logs show a legitimate user running a legitimate tool
Admiral Ackbar from Return of the Jedi saying it's a trap

Google, Chrome (Bawa), October 20, 2026.

Adversarial Design

Usability

Can people use it?

Adversarial Design

Can they use it toward someone else's goal?

Straight, “Adversarial intentionality,” Techné (accepted 2026).

Detector False Positives

Human-written essays flagged as AI  per 100, rounded
5.1%
US eighth-graders
61.3%
Non-native English (TOEFL)

Liang et al., Patterns 4(7), 2023. Average across seven detectors.

Evading Detection

70%
"Elevate the provided text by employing literary language."
→
Homer Simpson backing into a hedge until he disappears
3.3%
AI-essay detection after one rewrite prompt

Liang et al. 2023, Figure 2. 31 essays; proof of concept.

Training as a Control

Lain, Kostiainen and Čapkun, IEEE S&P 2022. 14,733 employees at one company, 15 months, before generative AI.

Expert Deference to an AI Agent

86%
of the time, experts went with the agent’s hidden lean, between two defensible options
Chance: 50%
0%100%
15%
Only 15% mentioned the agent

Herdağdelen, Joren and Tamkin, Anthropic, October 27, 2026. 50 practitioners; leans strengthened by design.

Sign-Off as Evidence

“Institutions often treat an expert’s sign-off as evidence of independent judgment, and as AI does more of the work, that assumption becomes shakier.”

Cartoon cat paw stamping a document that reads meow meow meow

Herdağdelen, Joren and Tamkin, Anthropic, 2026 (discussion).

Self-Attestation

What people did with AI and what they recall doing can differ

The interaction record
AI
Person
AI
Person
AI
Person
is stronger evidence than
“I reviewed it independently”

Straight, “So the Detectors Are Out,” ryanstraight.com, 2026.

Testing the Process

1
Was AI used?
2
Who chose, among which alternatives?
3
Is the interaction record kept?

Herdağdelen et al. 2026; Inside Higher Ed, August 5, 2026.

Out-of-Band Verification

Video calls can be faked end to end Requester Approver Payment ? Was the callback made before the money moved? Confirm on a channel the requester didn’t choose

Arup case: CNN, 2024; Hong Kong police via The Star, 2024.

Tool Miss Rates

“The AI flagged nothing”

is a claim to test

  • False-negative rate?
  • Measured on whose data?
  • Last re-tested when?

Cartoon dog sitting calmly at a table in a burning room

Liang et al. 2023; Heiding et al. 2026; Google, 2026; Cisco Talos, 2026.

Questions for the Business

Payments
When was the callback check last tested?
Reviews
Which decisions need a person’s own judgment?
Detection
What are the miss rates, on what data?
Awareness
Clicks after training, or only completion?
Exposure
What would an AI rank as worth selling?

Framework mappings for all five are on the handout.

The Binance Theft

1A fake recruiter and a malware “assignment”
2An internal AI tool used to reach other engineers
3One deployment token, then drained hot wallets
US$6.34 billion

Binance, initial findings, November 3, 2026.

Arup Revisited

Detection
No breach to detect
Training
He suspected a phish; the call overrode it
Out-of-band check
Found it, a week too late
A callback enforced before the first transfer would likely have stopped it.

CNN, 2024; Hong Kong police via The Star, 2024.

Questions

Slides, handout, sources and tools

ryanstraight.com

Ryan Straight · ryanstraight@arizona.edu

QR code linking to the talk page on ryanstraight.com