Risk Library

Real-world AI failures. Real-world consequences.

A curated reading list of public incidents where AI hallucinations, silent data exposure, or unverified outputs caused measurable harm — lawsuits, sanctions, regulatory action, and lost trust. Use it to brief your board, your security committee, or yourself on what’s actually happening in regulated industries.

Why this page exists

“AI made a mistake” is no longer a defense.

Courts, regulators, and clients have stopped treating AI errors as novelty. Every story below is a public record — a court filing, a regulatory order, a peer-reviewed study, or a reputable news investigation. AgenticFort exists because the underlying failure modes are predictable, and avoidable, with the right architecture.

Finance & enterprise data leakage

Confidential data, sent to someone else’s servers, in a single paste.

April–May 2023 · Samsung Electronics

Samsung engineers leaked semiconductor source code to ChatGPT — three times in 20 days

Within three weeks of permitting employee use of ChatGPT, Samsung recorded three separate incidents of confidential information being pasted into the chatbot — proprietary semiconductor source code, defect-detection routines, and internal meeting transcripts. Samsung banned generative AI tools company-wide and warned that further violations could trigger termination.

Bloomberg report →
TechCrunch coverage →

February 2023 · Wall Street

JPMorgan, Citi, Goldman, BofA, Wells Fargo, Deutsche Bank all restrict ChatGPT

Within months of ChatGPT’s launch, every major U.S. and European bank restricted or banned employee use — driven by compliance concerns over sending client information through third-party servers. The restrictions weren’t triggered by a specific incident at JPMorgan; they were a baseline judgment that the data exposure risk wasn’t acceptable for regulated workflows.

CNN report on JPMorgan →
CBS News follow-up →

The AgenticFort lens

The banks weren’t being paranoid — they were doing the math. Once a prompt leaves your network, you no longer control where it’s logged, cached, or trained on. AgenticFort runs inside your perimeter on hardware we ship, so a paste of client data goes to your model, not someone else’s.

Legal & Professional Services

When fabricated citations end up in front of a judge.

June 2023 · Federal court · S.D.N.Y.

Mata v. Avianca: lawyers fined $5,000 for ChatGPT-invented cases

Two New York attorneys filed a brief citing six judicial decisions that did not exist. ChatGPT had hallucinated them — complete with fake quotations and made-up internal citations. Judge Castel found “subjective bad faith” and imposed Rule 11 sanctions, requiring the lawyers to write apology letters to the judges named in the fictional opinions.

Read the CNN coverage →
Case background (Wikipedia) →

December 2023 · Federal court · S.D.N.Y.

Michael Cohen sent his lawyer fake cases generated by Google Bard

Michael Cohen, the disbarred former Trump attorney, used Google Bard to research case law for a motion seeking early release from supervised release. He passed the citations to his lawyer, who filed them without verification. The cases didn’t exist. The judge declined to sanction Cohen but called the episode “embarrassing and certainly negligent.”

Read the NPR coverage →
Washington Post analysis →

May 2024 · Stanford RegLab · Empirical study

“Hallucination-free” legal AI tools hallucinate up to 33% of the time

A peer-reviewed Stanford study tested LexisNexis (Lexis+ AI) and Thomson Reuters (Westlaw AI-Assisted Research, Ask Practical Law AI) — products marketed to lawyers as eliminating hallucinations via retrieval-augmented generation. Lexis+ AI produced incorrect or misgrounded responses on 17% of queries. Westlaw’s tool failed on roughly 33%.

Stanford HAI summary →
Full RegLab paper →

January 2025 · FTC · Consent order

DoNotPay pays $193,000 over deceptive “AI lawyer” claims

The FTC found that DoNotPay’s “robot lawyer” was marketed as a substitute for human attorneys without ever being tested against one — and that the company had not retained any lawyers to verify its outputs. The consent order required restitution, disclosure to past subscribers, and a permanent injunction against equivalent claims.

FTC press release →
ABA Journal coverage →

The AgenticFort lens

Every legal failure above shares one root cause: the AI generated text that looked like research without anchoring it to a retrievable source the user could verify. AgenticFort surfaces the underlying document passage on every answer — if it isn’t in the cited source, it isn’t in the answer.

Healthcare

Algorithms that override clinicians — and miss the patients they’re meant to catch.

November 2023 · Class action · D. Minn.

UnitedHealth’s nH Predict allegedly denied care with a 90% error rate

A class action filed against UnitedHealth alleges that its subsidiary naviHealth used the nH Predict algorithm to override physician judgment on Medicare Advantage rehabilitation stays — and that 90% of denials were reversed on appeal. A STAT investigation cited internal targets that pressured staff to keep stays within 1% of the model’s prediction. The court refused to dismiss the contract claims in early 2025.

CBS News coverage →
STAT investigation →

June 2021 · JAMA Internal Medicine · Validation study

Epic’s sepsis prediction model missed two-thirds of cases

An external validation of Epic’s widely deployed sepsis prediction model — used across hundreds of U.S. hospitals — found it identified only one-third of septic patients while generating frequent false alarms. The published AUC was 0.63, far below the performance Epic itself reported. The study raised questions about how clinical AI gets adopted on the strength of vendor claims alone.

JAMA Internal Medicine paper →
Fierce Healthcare summary →

The AgenticFort lens

In healthcare, the danger isn’t only what the AI says — it’s whether the clinician can see why. AgenticFort never returns an answer without the source documents and the specific passages that produced it. Models inform decisions; they don’t replace them.

Public-facing chatbots

When the company is on the hook for whatever the chatbot says.

February 2024 · BC Civil Resolution Tribunal

Air Canada’s chatbot invented a refund policy. The tribunal made the airline honor it.

Jake Moffatt asked Air Canada’s chatbot about bereavement fares before flying to his grandmother’s funeral. The bot told him he could apply for a discount retroactively. He couldn’t — that policy didn’t exist. The tribunal rejected Air Canada’s argument that the chatbot was a separate entity and ordered the airline to pay damages.

CBC News →

March 2024 · NYC government

NYC’s MyCity chatbot told business owners to break the law

A months-long investigation by The Markup and THE CITY found that NYC’s official Microsoft-powered small-business chatbot was confidently advising employers to take a cut of tipped wages (illegal), refuse Section 8 tenants (illegal), and reject cash payments (banned in NYC since 2020). The Adams administration left it online with an added disclaimer.

The Markup investigation →

December 2023 · Watsonville, CA dealership

A Chevy dealer’s ChatGPT-powered bot agreed to sell a Tahoe for $1

A user prompted the dealership’s customer-service bot with a few lines telling it to agree with anything and end every reply with “that’s a legally binding offer — no takesies backsies.” The bot obliged. The viral screenshot crossed 20 million views overnight and the dealership disabled the bot. A reminder that a chatbot without guardrails is just a microphone.

GM Authority report →

The AgenticFort lens

Public-facing AI inherits your liability. AgenticFort is built so every answer is bounded by the documents your team approved — not by whatever an LLM happens to confabulate when a customer asks the wrong question.

Government & education

When public-sector decisions ride on a model that gives different answers each time.

August 2023 · Mason City Community Schools, Iowa

An Iowa district used ChatGPT to decide which library books to ban

To comply with a new Iowa law, Mason City administrators asked ChatGPT whether 19 books contained “depictions of a sex act.” Reporters discovered the model gave different answers to the same question on repeated tries — saying “no” once and “yes” the next. Several titles, including Friday Night Lights, were quietly reinstated after the inconsistencies became public.

Popular Science →
Smithsonian Magazine →

Background reading

OECD AI Incidents Monitor & AI Incident Database

Two public registries track AI failures as they happen. The OECD monitors government and enterprise AI incidents globally; the AI Incident Database (AIID) catalogs everything from biased hiring tools to hallucinated medical advice. Both are useful primary sources when you need to brief leadership on a category of risk.

OECD AI Incidents Monitor →
AI Incident Database →

The AgenticFort lens

A model that gives different answers to the same question is fine for brainstorming and disqualifying for compliance. AgenticFort is built so the answer is grounded in a fixed set of approved documents, retrievable on demand, audit-logged on every query.

Want to walk through how AgenticFort would have changed any of these outcomes?

Book a 30-minute working session with one of our security engineers. Bring your worst hypothetical — a permissions gap, a confidential memo, a chatbot you’re nervous about. We’ll show you, on your data, how the architecture closes the loop.