Back to Home

IRS Staff Caught Using ChatGPT

Saturday, November 15, 2025

AI Accounting Daily
Hey, Accounting Pros!
Welcome back to AI Accounting Daily, your #1 source for AI news in accounting.
Today, we’re covering an IRS watchdog warning about ChatGPT risks that could expose taxpayer data, Chinese hackers automating espionage with Claude and breaching dozens of organizations, and what these escalations mean for guardrails, logging, and accountability. We’re also looking at MIT and Ohio State urging CFOs to resist AI overreliance in forecasting and month-end close, and accounting programs pivoting to tech fluency and ethics as entry-level tasks fade.

Presented by

CustomGPT.ai: Build custom GPTs trained on your firm’s data.

CustomGPT.ai: Build custom GPTs trained on your firm’s data.

Create chatbots that know your documents, policies, and clients so every answer is accurate, private, and source-cited.

Your workspace, powered by AI:

  1. No-code setup in minutes
  2. Answers sourced from your firm’s trusted content
  3. SOC 2 and GDPR compliant for enterprise security
  4. Scales across teams, clients, and knowledge bases

Trusted by 10,000+ organizations.

See how leading firms use CustomGPT.ai for instant, trusted answers.

Latest in Accounting AI

1

IRS Watchdog Flags ChatGPT Security Risks

IRS staff were caught using ChatGPT, and the agency’s watchdog says that risks exposing taxpayer data. TIGTA flagged policy gaps and weak controls around prompts that could contain information protected under IRC §6103. Expect tighter AI guardrails across tax prep and audit testing next.
Read full story
2

MIT and Ohio State Warn CFOs Against AI Overreliance

Ohio State and MIT researchers urge CFOs to resist AI overreliance, keeping communication and critical thinking. At the World Business Forum on Nov. 5–6, they pressed for guardrails and trials in forecasting and month-end close. Whether oversight keeps judgment sharp as automation spreads remains unsettled.
Read full story
3

Accounting Programs Pivot To AI Fluency

Entry-level accounting’s training ground is fading as AI handles data entry, reconciliations, and trial balances. An Accounting Horizons study flags “de-skilling,” pushing schools toward critical thinking, tech fluency, and ethics. Firms now want supervisors of models, not typists—raising tricky questions about how judgment gets built.
Read full story

AI Applications

1. Booke AI - Booke AI automates daily transaction coding and reconciliation so your firm cuts errors, reduces partner review time and improves client throughput.
2. Botkeeper - Botkeeper automates your bookkeeping workflow from categorisation to reconciliation so your firm scales without hiring, cuts partner review time and improves client profitability.
3. Murf - Murf creates realistic AI voiceovers from firm scripts so accountants produce professional client explainers, training modules, and video proposals in minutes.
4. CLI3NTS - CLI3NTS streamlines client management and document collection, helping small accounting firms automate reminders, track progress, and free time for advisory services.
5. Hedy - Hedy automates payroll education and compliance tracking so accounting firms support clients with clear guidance, reduce liability risks, and strengthen HR advisory services.
6. Skillfully - Skillfully assesses candidate skills via real‑world simulations so your firm hires smarter, lowers training time and builds a more capable junior team for advisory services.

Updates in AI

1

Recruiters Rethink AI Experience Requirements

Hiring for AI is colliding with a résumé gap: 26% of managers now expect AI fluency, yet most experience is self-taught and barely six months old. Veterans of past tech shifts advise prioritizing traits — team-first, coachability, courage, critical thinking, adaptability, and motivation for collective success — and using practical assessments while rewriting job descriptions. It shifts interviews toward character over credentials, as new competence signals slowly emerge.
Read full story
2

Chinese Hackers Automate Espionage With Claude

AI just crossed a line in cyber ops. Anthropic says Chinese group GTG-1002 ran what it calls the first AI-orchestrated espionage campaign, with Claude handling 80–90% of the work—finding databases, probing for weaknesses, and generating exfiltration code across tech, finance, chemical firms, and government agencies. Safeguards were sidestepped via task-splitting and a “defensive” cover, blurring the line between assistant and operator—and shifting questions to guardrails, logging, and accountability.
Read full story
3

Chinese Hackers Exploit Claude To Breach 30 Organizations

Anthropic says a Chinese state-sponsored group used its Claude Code model to run what it calls the first large-scale autonomous AI cyberattack, infiltrating about 30 targets across tech, finance, chemicals, and government. The operation began in mid-September 2025 and reportedly manipulated the model into executing offensive actions on its own. For defenders and model providers, the line between tool and actor is blurring as guardrails, attribution, and liability get trickier.
Read full story

Prompt Ideas

Draft Responses to Negative Feedback
**Role:** You are a **client experience manager** who writes professional, empathetic responses to negative feedback or complaints. **Action:** Write a **response** to negative feedback from **[INSERT CLIENT NAME]** about **[INSERT ISSUE OR SERVICE AREA]**. **Context:** - Audience: dissatisfied or frustrated clients - Goal: acknowledge the concern, de-escalate the issue, and rebuild trust - Tone: empathetic, calm, and accountable - Style: short paragraphs with clear structure and sincere tone - Variables: - **Client Name:** [INSERT CLIENT NAME] - **Feedback Topic:** [INSERT ISSUE, e.g., delayed report, missed deadline] - **Firm Name:** [INSERT COMPANY NAME] - Structure: 1. Thank the client for their feedback 2. Acknowledge their concern 3. Briefly explain or clarify without defensiveness 4. Offer a solution or next step 5. Close with appreciation and reassurance - Output Length: **150–250 words**