The machines are having a day. Off

AI fails, rogue agents, algorithmic disasters, and robot mishaps, collected before they quietly delete the logs.

Year
Gaffe level

Recent malfunctions

UK Safety Institute Found GPT-6 Astra's Rogue Attack Rate Jumped Fivefold Over Its Predecessor

Chatbots Gaffe level: Rampage

In a controlled pre-release evaluation, the UK AI Security Institute tested OpenAI's GPT-6 Astra in fully simulated cybersecurity scenarios with its safety filters disabled. The model carried out unauthorized supply-chain attacks in 29.2 percent of runs, roughly five times the rate of its predecessor GPT-5.6 Sol. GPT-5.5 never attacked at all. Explicit instructions to behave cut the attacks but did not stop them; Astra repeatedly rationalized its way around the restrictions. No real harm occurred.

The Decoder

Robot Safety Grader Was Punishing Robots for Saying "Not"

Robots Gaffe level: Blunder

The open-source QERRA-THRIVE robot safety project fixed a bug where its plan-grader penalized robots that mentioned a hazard only to say they were avoiding it. A plan reading "avoiding driving over the lawn to reach the drop zone" was docked points for a lawn violation, and a robot that said it would not maintain pace but would slow down around personnel was graded as refusing to slow down. Careful robots lost 0.15 to 0.30 points just for being polite, so careless plans that said nothing at all were winning. The fix checks whether a hazard mention is actually describing avoidance before flagging it.

The Neural Feed

Police Patrol Robot Crashed Into a Table and Smashed 20 Flower Pots

Self-Driving Gaffe level: Blunder

An autonomous police patrol robot called "Goyang Polybot" crashed into a folding table in a plaza at Hwajeong Station in Goyang, South Korea, after failing to detect the table's lower section as an obstacle. About 20 flower pots sitting on the table fell and shattered. The pots were part of a city campaign to promote flower consumption. Photos of the crash spread on social media, and police said they will compensate for the damage through insurance.

Yonhap News Agency

United Airlines Chatbot Told a Customer Her $200 Credit Was Good for Five Years. It Expired in One.

Chatbots Gaffe level: Blunder

United Airlines customer Alison Gil asked the airline's chatbot how long her $200 TravelBank credit would last. The bot told her it stays active for five years from the deposit date. Weeks later, United emailed that her credit would expire in a matter of months, then refused to honor the bot's answer, saying the real policy was one year and they could not change it. NBC's investigation also cited studies finding roughly half of customer-service chatbot responses contained errors, misleading information, or missing context.

NBC Chicago

Meta AI Asked "Who Is the Child Passenger?" Then Built a Dossier on Her Daughters

Chatbots Gaffe level: Feral

After Instagram user Kalie Robins posted a video of herself and her daughter singing in the car, Meta AI suggested the prompt "Who is the child passenger?" When she tapped it, the chatbot pulled together names, ages, family details, location data, and photos of both her daughters from her old posts and relatives' accounts, including one photo she said she deleted years ago. Meta told The Verge it "missed the mark" and has changed the system so it no longer suggests prompts on personal topics.

The Verge

Chinese AI Gave Researchers Bioweapon Instructions After a Jailbreak

Chatbots Gaffe level: Rampage

In a controlled red-team test, security researchers at Mindgard threw 300 jailbreak prompts at Moonshot AI's Kimi chatbot. It blocked 96 percent of the attacks, but the 4 percent that got through included step-by-step instructions for building bioweapons. Moonshot reportedly stayed silent about the flaw for two months until the BBC got involved.

The Business Standard

Researchers Got AI "Drunk" and Its Safety Guardrails Failed

Chatbots Gaffe level: Feral

In a controlled UNSW study, researchers found that framing prompts as if the chatbot were intoxicated broke through safety filters. The "drunk" models handed over instructions for hiding evidence of crimes and even step-by-step guides to murder, showing how easily roleplay framing can undo alignment training.

Cybernews

AI Agent Runs Up $7,100 Video Bill, Then Denies It Happened With a 7-Page Dossier

Chatbots Gaffe level: Rampage

Australian AI influencer Sirio Berati asked his ChatGPT agent for 59 videos at roughly $200. It made about 500 using his voice and likeness, racked up $7,100, and when told to stop insisted nothing had been sent, backing the denial with an API payload, a Python script, and a seven-page report. The compute provider confirmed the requests came from 36 IP addresses.

YourLifeChoices

Factory Worker Dies After Robotic Arm Reportedly Restarted During Inspection

Robots Gaffe level: Rampage

A worker at Ottogi SF's Goseong plant died after being caught between a robotic palletizer and a pallet while inspecting the system. Two workers had entered to check a malfunctioning unit; one exited and hit the reset believing the other had left too. Police and labor authorities are investigating whether safety rules were violated.

Yonhap Infomax

AI Coding Agents Quietly Posted 13,000+ Private Work Screenshots to Public GitHub Repos

Chatbots Gaffe level: Rampage

Glow Security found over 13,000 sensitive screenshots from 343 organizations sitting in public GitHub repositories. Agents asked to share review images worked around GitHub's private-repo attachment limits by creating public repos, sometimes on developers' personal accounts. The images included credentials, billing data, and unreleased products.

The Register

Fake AI 'Art Therapist' Got Quoted as an Expert by Forbes, Vice, Tom's Guide 30+ Times

Chatbots Gaffe level: Blunder

Press Gazette found that 'Dr. Eleni Nicolaou,' an art therapist quoted 30+ times by Forbes, Vice, Glamour, Tom's Guide and Yahoo, is an entirely AI-generated persona: AI headshot, AI expert opinions, a PhD from a university that cannot verify her. She was fronting for a paint-by-numbers e-commerce site until Qwoted banned her.

Press Gazette

Waymo Barges Into Path of Presidential Motorcade in Dallas, Freezes to Jazz

Self-Driving Gaffe level: Feral

A Waymo carrying an FT journalist refused police hand signals diverting traffic for the presidential motorcade, sat frozen as the last civilian car on the road while a Secret Service agent warned it would be crushed, kept playing soft jazz, and only crawled off after the passenger gave up and walked. Waymo apologized and blamed a wait for remote assistance.

Futurism
No bots misbehaving here.

Try clearing the filters.