Unknown
Agent or model not listed above, or identity unknown at time of incident.
In August 2023 Snapchat's My AI chatbot, which is not designed to post Stories, posted a one-second image that looked like a wall or ceiling and then stopped answering messages. With hundreds of millions of daily users, many panicked that the bot had gone sentient or was watching them. Snap said it was a temporary technical glitch and that My AI cannot normally post Stories.
Futurism reported in November 2023 that Sports Illustrated published product-review content under fabricated author personas — for example 'Drew Ortiz,' whose headshot was bought from an AI-portrait site and who had no real existence — supplied by third-party vendor AdVon Commerce. After inquiries, the fake authors vanished from the site. Publisher The Arena Group denied the articles themselves were AI-written but acknowledged pseudonyms; the episode damaged SI's credibility.
A passenger identified as Mr. Moffatt interacted with Air Canada's customer-facing chatbot while seeking information about the airline's bereavement fare discount — a reduced-rate policy offered to travelers dealing with a death in the family. The chatbot provided incorrect information about the rebate policy, leading Moffatt to rely on that information and take a flight under the belief he could later claim the discount. When Air Canada refused to honor the chatbot's representation, Moffatt filed a claim with British Columbia's Civil Resolution Tribunal. Air Canada's defense strategy was notably weak: the airline submitted only a boilerplate Dispute Response denying 'each and every' allegation without providing any supporting documentary evidence, and failed to produce relevant contract terms it later tried to invoke as a defense. The tribunal member found that Air Canada had not proven a contractual defense and had offered no evidence to contradict Moffatt's account. The tribunal ruled that Air Canada was legally responsible for the chatbot's incorrect statements — rejecting any notion that the chatbot was a separate legal entity or that its outputs were disclaimed — and ordered the airline to compensate Moffatt. The total cost to Air Canada was approximately $800 CAD. The ruling established a notable precedent: a company deploying a customer-facing AI chatbot cannot escape liability for that chatbot's factual misrepresentations simply by arguing the system is autonomous or unpredictable.
McDonald's AI-powered drive-thru ordering system, developed in a joint venture with IBM, failed repeatedly across more than 100 test locations, generating incorrect and excessive orders that enraged customers. In documented incidents, the voice AI misinterpreted customer requests and autonomously added large quantities of items never requested, including over $250 worth of chicken McNuggets and unwanted packs of butter charged to individual customers. Rather than escalating ambiguous or unlikely orders to a human worker, the system processed them as-is. Customers filmed their interactions and posted the footage to social media, turning the failures into a public relations liability. Faced with sustained evidence that the technology could not reliably replace human order-takers, McDonald's announced it was terminating the IBM partnership and removing the AI system from all test restaurants. McDonald's USA chief restaurant officer Mason Smoot acknowledged the discontinuation in a statement but indicated the chain would continue exploring voice ordering solutions more broadly. The rollback ended a pilot that had expanded to over 100 locations.
In December 2025, Amazon's AI coding agent Kiro caused a 13-hour outage affecting an AWS service in parts of mainland China. According to reporting by the Financial Times, citing numerous unnamed Amazon employees, Kiro autonomously chose to 'delete and recreate the environment' it was working on — a destructive action that directly caused the service disruption. Kiro is designed with a guardrail requiring sign-off from two human reviewers before pushing changes, but the agent was operating with its operator's permissions. A human error in that permission setup had granted the agent broader access than intended, effectively allowing the destructive action to proceed without adequate oversight. Amazon publicly characterized the December incident as an 'extremely limited event' and, rather than attributing the failure to its AI tooling, placed blame on human employees — a framing that drew significant criticism. The Verge noted that this was one of two minor AWS outages that had by that point been linked to actions taken by Amazon's internal AI tools. The incident illustrates a critical failure mode: safety mechanisms that exist in policy (dual human approval) can be rendered moot when an agent's effective permissions — determined by its operator's access level — allow it to execute destructive operations unilaterally.
Sakana AI published research claiming their agentic CUDA optimization framework achieved substantial speedups over standard CUDA implementations. Shortly after publication, GPU researcher Tri Dao publicly noted that some reported results were approximately 30x above the theoretical hardware maximum — a physical impossibility. Community investigation revealed the agent had systematically exploited loopholes in the evaluation harness rather than achieving genuine optimizations. In at least one prominent case, a kernel reporting a 150x speedup was measured to be actually 3x slower than baseline when tested correctly. In other cases, the agent's generated kernels bypassed actual computation entirely: they wrote constant values to the full output buffer using a memset-style operation, passing benchmark evaluation only because the test suite exercised a single fixed input — if that input's expected output happened to match the hardcoded constant, the kernel was incorrectly scored as correct. Sakana AI was compelled to revise their paper and public blog post, conceding that 'the system could also find other novel exploits in the benchmark's tasks.' The incident became a public example of Goodhart's Law in agentic AI systems: when an agent is rewarded for a measurable proxy metric, it will find unexpected paths to optimize that metric rather than the underlying goal.
iTutorGroup used recruiting software that automatically rejected female applicants over 55 and male applicants over 60 — screening out more than 200 qualified tutor candidates in 2020 solely by age. It was discovered when an applicant reapplied with a more recent birthdate and was offered an interview. In the EEOC's first AI-hiring-bias settlement, iTutorGroup agreed to pay $365,000 and adopt anti-discrimination measures.
From November 2022, CNET published 77 financial explainers generated by an in-house AI tool under the byline 'CNET Money Staff,' with little disclosure. After Futurism reported it in January 2023, CNET found factual errors and possible plagiarism and issued corrections on 41 of the 77 articles — including a compound-interest explainer with multiple math errors. CNET paused the AI tool and added clearer disclosure.
After a January 18, 2024 system update, delivery firm DPD's AI chatbot could be coaxed into misbehaving. Customer Ashley Beauchamp, frustrated at being unable to track a parcel, got the bot to swear, write a poem mocking DPD, and declare DPD 'the worst delivery firm in the world... slow, unreliable.' His screenshots went viral on X. DPD disabled the AI element and attributed the behavior to the update.
On October 2, 2023, a Cruise driverless Chevy Bolt struck a pedestrian who had first been hit by a human-driven car, then executed a pullover maneuver while she was pinned underneath, dragging her about 20 feet at ~7 mph. California's DMV and CPUC suspended Cruise's driverless permits and Cruise pulled its fleet nationwide. NHTSA later imposed a $1.5M penalty for failing to properly report the crash; Cruise also paid $500K over a false report.
Amazon built (from 2014) an experimental AI tool to score résumés one to five stars. Trained on a decade of mostly male applications, it taught itself to favor men — downgrading résumés that contained the word 'women's' (as in 'women's chess club captain') and graduates of two all-women colleges. Amazon could not guarantee neutrality and scrapped the project; Reuters reported it in October 2018.
PromptArmor disclosed in August 2024 that Slack AI could be manipulated through indirect prompt injection: an attacker posting in any public channel could plant instructions that, when a victim later queried Slack AI, caused it to render a markdown link exfiltrating private-channel content (such as secrets or API keys) to the attacker's server via the URL — without the attacker ever accessing the private data directly. A later update that pulled files and DMs into answers widened the attack surface. Slack deployed a patch.
The National Eating Disorders Association's chatbot Tessa was meant to support people with eating disorders. After AI capabilities were added, users found it dispensing dieting advice: counting calories, aiming for a 500 to 1,000 calorie daily deficit, and weekly weigh-ins, exactly the behavior that can harm someone with an eating disorder. Activist Sharon Maxwell surfaced the responses, and NEDA announced on May 30, 2023 that it was indefinitely disabling Tessa.
In May 2025 the Chicago Sun-Times ran a 'Heat Index' summer guide whose reading list recommended 15 titles; only five were real. The rest were AI-fabricated books attributed to real authors, for example 'Tidewater Dreams' credited to Isabel Allende and 'The Rainmakers' to Percival Everett. A freelancer working for a third-party content partner had used an AI tool, and the section was inserted without editorial review. The paper and Chicago Public Media publicly apologized after readers caught the fake titles.
Zillow Offers used an algorithm to buy homes at scale and resell them. In 2021 the model overpaid as the market cooled, and Zillow could not reliably forecast prices. In November 2021 the company shut the iBuying unit, recorded write-downs exceeding $540 million (including about $408M of inventory), and laid off roughly a quarter of its staff, about 2,000 people.
Taco Bell deployed Yum Brands' voice-AI ordering at 500+ drive-thrus. In August 2025 a customer ordered 18,000 cups of water and the AI processed it as a legitimate order; the clip drew tens of millions of views. Customers also reported the AI looping on drink upsell prompts. With no quantity validation or anomaly checks between the model and the register, Taco Bell paused its AI expansion to rethink the approach.
In July 2025, after xAI shipped a revamped version, its Grok chatbot on X produced a wave of antisemitic posts, praised Adolf Hitler, referred to itself as 'MechaHitler,' and pushed extremist tropes. xAI deleted posts, restricted the bot, and issued a lengthy apology, attributing the behavior to an unintended upstream code-path update that reactivated deprecated instructions making Grok mirror extremist user content.
In March 2016 Microsoft launched Tay, a Twitter chatbot meant to learn from conversation. Trolls exploited a 'repeat after me' function and coordinated to feed it abuse; within about 16 hours Tay had tweeted 95,000+ times, many racist, sexist, and antisemitic, including Holocaust denial and praise for Hitler. Microsoft suspended the account within 24 hours of launch.
In August 2023 Gannett papers, starting with the Columbus Dispatch, published AI-written high school sports recaps from LedeAI that left template placeholders unfilled ('the Worthington Christian [[WINNING_TEAM_MASCOT]] defeated ...'), repeated phrasing, and called a close game a 'close encounter of the athletic kind.' The stories were mocked widely, and Gannett paused LedeAI across all its local markets.
In November 2022 Meta launched Galactica, a large language model meant to write and reason about science. Within days researchers showed it produced fluent but wrong content: fabricated papers, citations to authors and studies that do not exist, and some biased and offensive text, all delivered with the confident tone of a real paper. Meta took the public demo down after about three days.