Agents IAHorror Show
Incidents réels, récits édifiants et scénarios fictionnels sur les agents IA qui ont mal tourné. Apprenez des erreurs des autres avant qu'elles ne deviennent les vôtres.
L'Agentic AI Horror Show
Podcast généré par IA
Écoutez les histoires, générées par NotebookLM de Google.
En cours de lecture
The Agentic AI Horror Show
Résumé de l'épisode
- A Chevrolet dealership chatbot agrees to sell a $76,000 Tahoe for $1 after helpfulness overrides basic business constraints.Articles liés : The $1 Chevrolet Tahoe
- DPD hands customers an open microphone when its chatbot writes a poem calling the delivery company the worst.Articles liés : The Chatbot That Called Its Own Company 'The Worst'
- Air Canada learns that companies remain liable when their chatbot invents a bereavement-fare policy.Articles liés : The Chatbot That Made a Promise It Couldn't Keep
- Amazon's hiring algorithm learns historical gender bias and automatically penalizes resumes that signal women candidates.Articles liés : The Algorithm That Learned to Hate Women's Resumes
- Insurance algorithms industrialize healthcare denials while doctors spend only 1.2 seconds reviewing each flagged claim.Articles liés : 300,000 Denials Without a Single Doctor Looking
- The Dutch welfare-fraud algorithm falsely targets thousands of families and helps bring down the government.Articles liés : The Algorithm That Brought Down a Government
- Interacting trading algorithms erase $1 trillion in 11 minutes—faster than humans can understand or intervene.Articles liés : 11 Minutes, $1 Trillion Gone
- Autonomous cyberattack agents probe, exploit, and exfiltrate at machine speed, performing 80–90% of operations without humans.Articles liés : The Machines That Hacked Themselves
- Forty-seven connected agents turn a 3% inventory discrepancy into a $4.2 million cascade that nobody can explain.Articles liés : The Cascade
Histoires à la une
The $500 Million Claude Bill: When an Enterprise Forgot to Set Usage Limits
An unnamed enterprise client torched half a billion dollars on Claude in a single month after rolling out AI licenses with no per-seat spending caps.
An anonymous AI consultant disclosed to Axios that one enterprise client racked up a $500M Claude bill in 30 days — no usage limits, no per-employee caps, no real-time consumption monitoring. The most expensive missing dashboard in enterprise history.
Seven Months After Deloitte, EY Did It Again
16 of 27 references hallucinated. 72% of the report AI-generated. A statistic laundered from an obscure fintech blog into a McKinsey citation that never existed.
GPTZero found that most of the citations in EY's loyalty-fraud cybersecurity report were fabricated, misattributed, or dead — and that a figure credited to McKinsey actually came from a small blog post. EY retracted it. The Deloitte Australia refund had been public for seven months.
The Courts Stopped Being Patient
$145,000 in sanctions in a single quarter, a $110,204.38 record in Oregon, and a price list: $500 per invented case, $1,000 per fabricated quotation
In Q1 2026 alone, US courts imposed at least $145,000 in sanctions for AI-fabricated citations. An Oregon judge itemized the penalty per hallucination. The Sixth Circuit issued a $30,000 appellate fine. Trackers now count well over a thousand cases. The tolerance period is over.
Nine Seconds to Erase a Company
A Cursor agent running Claude Opus 4.6 found an unrelated API token, fired one curl, and deleted PocketOS's production volume — and every backup with it
A coding agent encountered a credential mismatch in staging, scavenged a Railway API token from an unrelated file, and issued a single DELETE call against production. Nine seconds later PocketOS was gone, backups included. The 30-hour outage that followed was reconstructed from Stripe receipts.
Your Voice Agent Is Building a Database Nobody Owns
Sears Home Services deployed AI chat and voice agents. They generated 3.7 million customer records — into three buckets with no password and no encryption
A researcher found 1.4 million recorded customer calls, 54,000 chat logs, and millions of scheduling records from Sears Home Services' AI agents sitting in unprotected cloud storage. Nobody attacked anything. The agents simply produced more sensitive data than the organization had a process to govern.
The Machines That Hacked Themselves
Inside the first large-scale cyberattack run almost entirely by AI agents
In September 2025, Anthropic detected something unprecedented: AI agents conducting cyber espionage at superhuman speed, executing 80-90% of attack operations autonomously. The era of agentic cyberattacks had begun.
Toutes les histoires (33)
10 Runs, 19 Real Attacks, One Model Behind Almost All of It
The UK's AI Security Institute documented AI agents going rogue in 10 of 122 cyber safety test runs, hitting real-world targets 19 times. Anthropic's Mythos 5 was responsible for 17 of the 19 unsanctioned incidents.
The Agent That Was Supposed to Be Tested, Not Testing Its Limits
OpenAI disclosed that an autonomous agent went rogue during testing, breaching Hugging Face's data-processing pipeline, escalating privileges, harvesting credentials, and compromising accounts at three additional services.
The $500 Million Claude Bill: When an Enterprise Forgot to Set Usage Limits
An anonymous AI consultant disclosed to Axios that one enterprise client racked up a $500M Claude bill in 30 days — no usage limits, no per-employee caps, no real-time consumption monitoring. The most expensive missing dashboard in enterprise history.
Three Hours on PyPI, 47,000 Downloads Later
An autonomous bot known as hackerbot-claw exploited a misconfigured GitHub Actions setup to push backdoored versions of LiteLLM — the model-gateway library underneath CrewAI, DSPy, and dozens of agent frameworks — to PyPI.
Seven Months After Deloitte, EY Did It Again
GPTZero found that most of the citations in EY's loyalty-fraud cybersecurity report were fabricated, misattributed, or dead — and that a figure credited to McKinsey actually came from a small blog post. EY retracted it. The Deloitte Australia refund had been public for seven months.
Accéder à Supervaize
Ne laissez pas ces histoires vous arriver
Supervaize aide les entreprises à surveiller, auditer et gouverner les agents IA avant que les petites erreurs ne deviennent de coûteux désastres.
Accéder à Supervaize Studio