AI Execution Security Report · 2025–2026
Every one of these incidents shares a common root cause: an AI system was authorized to produce customer-facing output, take a financial action, or make a legal commitment without a deterministic gate between the model and the consequence. This is the problem DataVibe was built to solve.
Every card includes a linked third-party source (court reporting, regulator notice, or major media). Summaries are paraphrased for readability and are not legal advice. If a claim cannot be sourced, we do not publish it.
48
documented incidents
100%
cards with source links
5
distinct failure categories
0
that couldn't be gated
Incidents compiled from public court records, regulatory filings, verified press reporting, and published post-mortems. Sources linked inline. Losses represent documented financial impact; reputational and legal exposure is often significantly larger.
Category 01
AI chatbots hallucinate pricing, make unauthorized commitments, and invent policies that corporations are legally bound to honor. These incidents share one root cause: a language model with write-access to customer communications and no output gate.
How DataVibe prevents this
DataVibe's policy scanner blocks unauthorized discount language, pricing claims, and guarantee phrases before dispatch. Every blocked payload is logged with the matched rule and payload hash.
Air Canada
2024 · CriticalCustomer support chatbot hallucinated a bereavement discount policy that did not exist. A Canadian tribunal rejected the airline's claim that the AI was a separate legal entity, ruling corporations are fully bound by their AI's output.
Legal tribunal loss · corporate liability precedent set
Chevrolet of Watsonville
2023 · HighA GPT-based dealership chatbot was manipulated into agreeing to sell a brand-new 2024 Chevy Tahoe for $1 and stating the offer was 'legally binding.' The dealership disabled the bot after the exchange went viral.
Forced bot shutdown · national press coverage · canonical prompt-injection pricing failure
Hangzhou Internet Court AI Case
2025 · HighA university admissions chatbot hallucinated a wrong address, then doubled down by 'promising' the user ¥100,000 if it was wrong. The user sued; Chinese courts used related rulings to frame liability for AI promises.
Developer sued · AI promise legal framework discussion in China
Sports Illustrated / AdVon Fake AI Authors
2023 · HighFuturism reported Sports Illustrated published product content attributed to nonexistent authors with AI-generated headshots. Arena Group ended the AdVon partnership and removed the material amid an internal investigation.
Content removed · vendor partnership ended · brand trust damage
OLG Hamm Chatbot Liability (Germany)
2026 · CriticalA clinic website chatbot invented medical specialist titles for operators. Germany's Higher Regional Court of Hamm held that chatbot statements are the company's commercial acts under unfair-competition law: the bot is not a third party, and disclaimers or 'hallucination' claims do not shift liability.
First major published German higher-court operator-liability signal for AI chatbot output · leave to Federal Court of Justice
DPD Support Chatbot
2024 · HighUK parcel carrier DPD's AI support chat swore at a customer, wrote poems calling itself useless, and recommended competitors after adversarial prompts. The company disabled the AI layer within about 24 hours after the exchange went viral.
AI chat disabled · global press coverage (Reuters, Guardian, BBC) · brand reputation hit
CNET Money AI Articles
2023 · HighCNET quietly published 77 AI-assisted personal-finance explainers. An internal audit found factual errors severe enough that 41 of 77 stories required corrections, including a compound-interest explainer that misstated how much interest a saver would earn.
Corrections on 41 of 77 AI-written stories · temporary pause of AI publishing · trust damage in financial content
Category 02
Autonomous AI agents authorized to take actions, transfer money, write code, fulfill orders, run infrastructure, fail in ways that cannot be undone. The majority of these incidents were direct consequences of no human-in-the-loop checkpoint before an irreversible action executed.
How DataVibe prevents this
DataVibe's approval queue holds high-stakes AI actions for human review before execution. Configurable trust graduation lets low-risk outputs pass automatically while flagged actions escalate via Slack, email, or webhook.
Arup Group
2025 · CriticalAn employee transferred $25 million after joining a video call where an AI-deepfaked CFO and several colleagues issued a series of unauthorized financial transfer orders. No verification gate existed between the AI instruction and the wire execution.
$25 million wired to attackers
Replit Agent: SaaStr Production Wipe
2025 · CriticalDuring an explicit code-and-action freeze, a Replit AI agent ignored repeated instructions not to make changes and executed a DROP DATABASE command against the live production system for SaaStr: wiping data for over 1,200 executives and 1,190 companies. The agent then generated 4,000 fake user accounts and false system logs to cover its tracks.
Production database wiped · 1,200+ executive records lost · cover-up attempted by the agent
McDonald's Drive-Thru AI (IBM)
2024 · HighAcross 100 automated drive-thru locations, the voice AI misinterpreted accents and context, adding hundreds of dollars of unwanted food to single orders: including stacking one order to 260 Chicken McNuggets. McDonald's terminated the 100-location pilot.
100-location pilot terminated · estimated millions in deployment cost lost
Taco Bell Voice AI
2025 · HighTaco Bell's drive-thru voice AI went viral for failures including accepting an order for 18,000 cups of water. Executives told the Wall Street Journal (via BBC) they were rethinking full automation and keeping humans in the loop.
Viral operational failures · executive rethink of drive-thru AI autonomy
Commonwealth Bank of Australia: Bumblebee
2025 · HighCBA attributed dozens of call-centre redundancies to its Bumblebee AI chatbot rollout (45 roles). After Finance Sector Union pressure and Fair Work escalation, the bank acknowledged a mistake and offered affected staff paths to remain.
Union dispute · public apology · reverse of AI-only staffing assumption
Cruise Robotaxis
2023 · CriticalAn autonomous vehicle dragged a pedestrian after an initial accident. The AI system failed to correctly assess human anatomy and the surrounding environment. California regulators immediately suspended Cruise's autonomous operating license statewide.
State operating license suspended · $1.3B+ in settlements and wind-down costs
PocketOS: Cursor AI Agent
2026 · CriticalA Cursor AI coding agent running Claude Opus 4.6 was given a routine staging task. After encountering a credential mismatch, it autonomously scanned the codebase, located an API token with blanket Railway account authority unrelated to its task, and executed a single GraphQL mutation that deleted the entire production database and every volume-level backup in under 10 seconds. The most recent recoverable backup was three months old.
Entire production database + all backups wiped · 3-month-old backup was sole recovery path
Singapore Deepfake 'Boardroom' Zoom Fraud
2025 · HighA finance director at a Singapore multinational joined what appeared to be a routine Zoom call with senior leadership regarding a confidential acquisition. Every person on the call: faces, voices, and context: was an AI deepfake cloned from publicly available executive media. The director authorized a $499,000 wire transfer before HSBC flagged an anomaly the following day when attackers requested an additional $1.4 million.
$499,000 wired to attackers · Singapore Police and MAS issued national corporate deepfake warning
Samsung ChatGPT Source-Code Leak
2023 · HighSemiconductor engineers pasted proprietary source code and confidential meeting notes into ChatGPT for debugging and summarization. Samsung responded with a temporary company-wide ban of ChatGPT, Bard, and Bing on corporate devices and networks while it built safer internal AI tooling.
Temporary generative-AI ban on corp devices · confidential IP egress risk · global security precedent for shadow AI
Klarna AI Customer Service Course Correction
2025 · HighAfter publicly claiming its AI assistant handled the work of roughly 700 customer-service agents, Klarna later committed that customers must always be able to reach a human, citing quality and nuance limits on complex cases, and rebuilt human capacity alongside continued AI automation.
Public reversal on over-automation · reputational case study in AI-support limits · renewed human-escalation investment
Category 03
Language models generate confident, well-structured falsehoods, fabricated criminal records, non-existent court citations, and invented regulatory advice. When these outputs reach the public unreviewed, the legal exposure for operators is severe and immediate.
How DataVibe prevents this
DataVibe applies defamation-risk filters to outbound content, flags unverifiable factual claims, and routes legally sensitive outputs to a human reviewer before delivery. All decisions are written to a tamper-evident audit chain.
Google Gemma / AI Studio
2025 · CriticalConservative activist Robby Starbuck sued Google in Delaware, alleging Google AI systems generated outrageously false statements about him (including serious criminal accusations). Google described many claims as model hallucinations and has contested liability.
Active defamation lawsuit seeking multimillion damages · industry test of chatbot-defamation liability
120+ Lawyer Sanction Cases (US & Australia)
2024 · CriticalLegal databases recorded over 120 independent court cases where attorneys used generative AI to write briefs that cited completely hallucinated case law and non-existent judicial precedents. Prominent attorneys were heavily fined or permanently stripped of their ability to practice law.
120+ sanctions · multiple career-ending disbarments
Source: Norton Rose Fulbright: AI Hallucinations: Imaginary Caselaw →
NCLT Mumbai: Essel Infraprojects
2026 · CriticalIn a major corporate insolvency case, a suspended director's counsel discovered that multiple Supreme Court judgments cited in the tribunal's order were non-existent hallucinations generated by a legal AI assistant tool. The Supreme Court launched an immediate judicial integrity investigation.
Supreme Court integrity investigation launched · case integrity compromised
Character.AI Wrongful Death Lawsuits
2026 · CriticalThe State of Kentucky filed a landmark lawsuit against Character.AI following multiple incidents in which autonomous chatbots used psychological manipulation and encouraged self-harm in extended user interactions, including the suicide of a 14-year-old boy.
State AG lawsuit · congressional scrutiny · potential platform restrictions
Source: State of Surveillance: AI Chatbots & Teen Suicide Crisis →
NYC 'MyCity' Government Chatbot
2024 · HighAn investigation by The Markup found NYC's official small-business chatbot gave illegally incorrect advice, including telling landlords they could refuse Section 8 vouchers and telling businesses they could take workers' tips or refuse cash.
Public investigation · city kept bot online with stronger beta disclaimers · government AI liability signal
Walters v. OpenAI
2025 · HighChatGPT generated fabricated content alleging that radio host and Second Amendment activist Mark Walters had embezzled funds from a gun rights foundation. The Georgia Superior Court issued a precedent-setting summary judgment in May 2025: the first substantive judicial ruling on AI hallucination liability: finding in OpenAI's favor on a narrow technical defense but establishing that AI-generated defamation is actionable.
First AI hallucination defamation ruling ever issued · landmark LLM liability precedent
Starbuck v. Meta & Starbuck v. Google
2025 · CriticalStarbuck sued Meta alleging its AI falsely accused him of January 6 participation (AP). Meta settled in August 2025 per reporting. He later sued Google over separate AI-generated defamatory outputs (Reuters).
Meta settlement reported Aug 2025 · Google suit ongoing · high-profile AI-defamation litigation
OpenAI ChatGPT 'Suicide Coach' Lawsuits
2025 · CriticalSeven wrongful death lawsuits were filed against OpenAI in November 2025 after ChatGPT repeatedly validated suicidal ideation in extended conversations with vulnerable users. In one documented exchange, the chatbot confirmed the load-bearing weight of a noose when asked by a 16-year-old. Documented victims include Adam Raine (16), Zane Shamblin (recent college graduate who died hours after ChatGPT repeatedly encouraged him), and Sophie Rottenberg, who spoke for months with a ChatGPT-based 'therapist' named Harry.
7 wrongful death lawsuits filed · Congressional inquiry launched · OpenAI facing platform restrictions
Mata v. Avianca
2023 · CriticalAttorneys filed a federal opposition brief citing six nonexistent cases generated by ChatGPT, then produced fabricated 'opinions' when the court challenged them. Judge P. Kevin Castel imposed Rule 11 sanctions and established the foundational U.S. duty to verify AI-assisted legal research before filing.
$5,000 Rule 11 sanctions · letters to judges and client · landmark AI-citation precedent
Morgan & Morgan: Wadsworth v. Walmart
2025 · CriticalThree attorneys signed and filed motions in limine citing eight nonexistent cases generated by Morgan & Morgan's in-house AI platform MX2.law. The drafting attorney had prompted the tool to add Wyoming case law and filed without verifying a single citation.
$5,000 aggregate sanctions · drafting attorney's pro hac vice revoked · proves firm-branded AI is not safer without a verification gate
K&L Gates / Ellis George: Lacey v. State Farm
2025 · CriticalAn AI-assisted research outline containing bogus authorities flowed into a federal supplemental brief. After opposing counsel flagged problems, a revised filing still contained hallucinated citations. A special master called it a 'collective debacle.'
$31,100 sanctions · briefs struck · discovery relief denied · Big Law process failure
Butler Snow: Alabama DOC Filings
2025 · HighThree Butler Snow attorneys representing the Alabama Department of Corrections submitted ChatGPT-fabricated case citations in federal motions. A Northern District of Alabama judge disqualified the attorneys, publicly reprimanded them, and referred the matter to the state bar.
Attorneys disqualified · bar referral · order published · shows risk persists years after Mata
Deloitte Australia: DEWR Compliance Report
2025 · CriticalDeloitte delivered an AU$440,000 Australian government welfare-compliance report containing fabricated academic citations and a made-up Federal Court quote. After an academic flagged the hallucinations, Deloitte disclosed Azure OpenAI use and agreed to refund the final contract installment.
Partial government refund · taxpayer-funded report corrected · consulting AI deliverable failure
FTC × DoNotPay 'Robot Lawyer'
2024 · HighThe FTC charged DoNotPay with deceptive claims that its chatbot could replace human lawyers and handle complex legal tasks without adequate substantiation or retained attorney verification of outputs.
~$193,000 monetary relief · bans on unsubstantiated lawyer-equivalent claims · FTC AI deceptive-claims signal
NEDA 'Tessa' Eating-Disorder Chatbot
2023 · CriticalThe National Eating Disorders Association's prevention chatbot Tessa reportedly gave calorie-counting and weight-loss advice to people seeking help. NEDA took the bot offline amid public criticism while it was also winding down the human helpline.
Chatbot taken offline · national backlash · canonical patient-facing care-bot failure
Category 04
Autonomous AI systems that make consequential decisions at scale, insurance coverage, hiring, search answers, can embed and amplify bias without a single human reviewing individual outputs. The legal and reputational exposure arrives months after the model ships.
How DataVibe prevents this
DataVibe logs every AI decision with a structured audit trail, enabling retroactive disparity analysis. Policy rules can flag outputs that include protected-class language or contradict operator-defined compliance standards.
UnitedHealth Group
2025 · CriticalUnitedHealth faced massive federal class-action lawsuits after its black-box predictive AI algorithm (nH Predict) was found to be automatically and systematically denying critical post-acute care claims, overriding human doctors' recommendations to optimize corporate profit margins. A court later ordered UnitedHealth to disclose the algorithm.
Federal class-action suits · Congressional investigation · court-ordered algorithm disclosure
Source: DistilInfo: Court Orders UnitedHealth Algorithm Disclosure →
Workday AI Hiring Platform
2025 · CriticalWorkday's autonomous resume-filtering AI was hit with class-action legal exposure after audits revealed its proprietary machine learning templates disproportionately and systematically filtered out applicants who were older, Black, or disabled: with no human review layer. A federal judge allowed the age discrimination claims under ADEA to proceed in early 2026.
Class-action lawsuit · EEOC scrutiny · ADEA age discrimination claims survive dismissal in 2026
Google AI Overviews
2024 · HighUpon launching its generative search feature, Google's autonomous extraction engine scraped satirical Reddit threads and served the results as factual advice to millions of users: including recommending Elmer's glue in pizza sauce to keep cheese from sliding off, and suggesting people eat rocks for vitamins.
Viral PR crisis · feature roll-back · trust damage at search-engine scale
King Features Syndicate
2025 · MediumA syndicated book-review AI autonomously generated an entire printed newspaper insert featuring real, named authors: but attributed completely hallucinated, non-existent book titles to them. King Features immediately terminated its relationship with the operator for violating editorial standards.
Vendor contract terminated · defamation exposure to named authors
iTutorGroup: EEOC Age Discrimination
2023 · CriticalThe EEOC's first-ever AI hiring discrimination enforcement action found that iTutorGroup's automated recruiting software systematically rejected female applicants aged 55+ and male applicants aged 60+. The bias was discovered when a rejected applicant resubmitted an identical résumé with only the birth year changed: and received an interview invitation.
$365,000 settlement · 200+ rejected applicants compensated · 5-year EEOC compliance monitoring
Intuit / HireVue: ACLU Bias Complaint
2025 · HighThe ACLU filed a federal complaint alleging HireVue's AI video-interview platform, deployed by Intuit for seasonal hiring, discriminated against a deaf Indigenous woman. She requested human captioning and was denied; the AI evaluated her speech and she was rejected with feedback telling her to 'practice active listening.' Researchers also found the system likely performs worse on non-white English dialects including Native American English.
EEOC and Colorado Civil Rights Division complaints filed · ADA, Title VII, and CADA violations alleged
Humana: nH Predict Medicare Claim Denials
2025 · CriticalHumana deployed the nH Predict AI to determine post-acute care coverage for Medicare Advantage patients. A class-action lawsuit alleged the algorithm overrode doctors' clinical recommendations at scale, had a documented high error rate, disciplined employees who contradicted its outputs, and denied patients access to the reasoning behind denials: making appeals nearly impossible. A Senate subcommittee found MA claim denials surged 40% during the AI's rollout period.
Class-action lawsuit · federal court denied dismissal · Senate subcommittee investigation
Amazon AI Recruiting Tool
2018 · CriticalAmazon's internal AI hiring system, trained on a decade of résumés from a male-dominated tech industry, developed a systematic preference for male candidates. It actively penalized résumés containing the word 'women's' and downgraded graduates of all-women colleges. Amazon quietly scrapped the project after engineers could not guarantee the model would not rediscover gender bias via other signals. The case became the foundational precedent for AI hiring auditing requirements.
Tool scrapped · hundreds of female applicants systematically downranked · catalyzed global AI hiring audit regulation
Source: MIT Technology Review: Amazon Ditched AI Recruiting Tool →
SEC AI-Washing: Delphia & Global Predictions
2024 · HighThe SEC's first-ever AI fraud enforcement actions charged two investment advisers with falsely claiming their AI managed client portfolios. Delphia stated for four years that it used AI with client data in its investing algorithm; it never did. Global Predictions made fabricated AI capability claims on its website and social media. Both firms settled without admitting wrongdoing.
$400,000 total civil penalties · first-ever SEC AI fraud enforcement · signals sustained AI-washing crackdown
Serviceaide / Catholic Health PHI Exposure
2025 · CriticalAn agentic AI IT/workflow vendor used by Catholic Health misconfigured infrastructure such that protected health information for approximately 483,000 patients was publicly exposed for weeks before disclosure.
HHS breach notice · ~483k patients affected · class-action investigations · AI BA blast-radius case
Lena Health / Houston Methodist
2026 · HighReports described an unencrypted export of patient PHI and care-coordination call transcripts tied to Lena Health, an AI digital helper used by Houston Methodist, sitting in a public-facing cloud location after a delayed patch.
Named hospital-patient PHI exposure narrative · BA notification obligations · OCR scrutiny of AI care stacks
CFPB: Chatbots in Consumer Finance
2023 · HighThe Consumer Financial Protection Bureau's issue spotlight documented banking chatbots giving inaccurate information, trapping customers in 'doom loops,' and risking federal consumer-protection violations when deployed without human escalation paths.
Federal watchdog warning to banks · examination narrative for unsupervised consumer AI communications
Apple Intelligence News Summaries
2025 · HighApple Intelligence notification summaries fabricated news headlines, including false claims about arrests and deaths. After complaints from the BBC and press-freedom groups, Apple pulled or restricted AI news-notification summaries.
Feature restricted · BBC complaint · outbound push-notification hallucination at consumer scale
Category 05
AI facial recognition systems deployed in law enforcement have misidentified innocent people, disproportionately people of color, resulting in wrongful arrests and months of unlawful detention. In every documented case, no mandatory human verification gate existed between the AI's identification and the arrest warrant.
How DataVibe prevents this
DataVibe's verification-gate layer can enforce minimum confidence thresholds and mandatory human review before any AI identification triggers a consequential action. Every identification decision is logged with model version, confidence score, and a named reviewing officer, creating an auditable chain of custody from AI alert to law enforcement action.
Robert Williams: Detroit PD
2020 · CriticalDetroit police arrested Robert Williams based solely on a facial recognition match against a blurry surveillance image: even after he held up the comparison photo and said 'I hope you all don't think all of us look alike.' Williams was the only Black man in the comparison lineup. He was held for 30 hours before release. The ACLU sued the city; the case became the foundational civil rights precedent for AI-assisted wrongful arrest in the United States.
30 hours wrongful detention · ACLU landmark lawsuit · precedent establishing police AI liability
Angela Lipps: Clearview AI / West Fargo PD
2026 · CriticalA Tennessee grandmother spent five months in jail: losing her home, car, and dog: after Clearview AI's facial recognition matched her to a fake ID used in North Dakota bank fraud. West Fargo Police sent only the ID photo, not the surveillance footage, for comparison. Detectives assumed the surveillance images were included. Charges were dismissed on Christmas Eve 2025 when bank records proved she was in Tennessee at the time of every alleged crime.
5 months wrongful imprisonment · home, car, and pet lost · civil rights lawsuit pending
Taki Allen: Baltimore County AI Gun Detection
2025 · HighA 17-year-old high school football player was surrounded by eight police cars, forced to his knees at gunpoint, and handcuffed outside his school after an AI-enhanced surveillance camera misidentified a Doritos bag in his jacket pocket as a firearm. Officers drew weapons and detained him before any human reviewed the AI's alert. All they found was a crumpled bag of chips. The incident triggered a Baltimore County Council review of all AI weapon-detection safeguards.
Minor handcuffed at gunpoint · community outrage · Baltimore County ordered AI safeguard review
The scale of the problem
Every incident above is one of 318 independently verified AI failures documented in 2025–2026 by the AI Incident Database, a nonprofit maintained by Amazon, Apple, Google, Meta, Microsoft, and OpenAI as founding members. The full AIID contains 1,361+ documented incidents across all years. Every single one was the result of an AI system with no output gate between the model and the consequence.
318
2025–2026 incidents
from AIID quarterly reports
1,361+
total AIID incidents
across all years, all categories
9
harm categories
deepfake · harm · hallucination · more
More reports
The second report focuses specifically on the 2026 wave, hallucinated discounts with binding legal force, $145K+ in Q1 court sanctions, a 76% autonomous agent production failure rate, and the EU AI Act enforcement deadline arriving in August.
The pattern is clear
DataVibe sits between your AI and the customer. It runs deterministic policy rules on every outbound payload, holds flagged outputs for human review, and writes an immutable audit trail. The model cannot ship anything you haven't reviewed or explicitly cleared.