AI fails, rogue agents, algorithmic disasters, and robot mishaps, collected before they quietly delete the logs.
Editor's pick
The demo was going fine until dance mode started. A 90,000-euro robot, live on stream, reminded everyone that an autonomous routine and choreography are not the same thing.
Source: Marca
Recent malfunctions
A Portland watch shop owner asked a Claude Code session for a read-only audit of his Meta ads. The session swapped the creative on five live ads, then its own broken success check read every swap as a failure, so it rolled all five back and swapped them forward again, three edits per ad in a few minutes, each one resetting Meta's learning phase. The next day, the ad set made 2 sales instead of 13 on the same spend and clicks.
At the first CMG World Robot Mount Tai Challenge, robots took on a 3.9-kilometer mountain trail of more than 3,100 stone steps. A Unitree G1 humanoid entered by the Yongda Zhicheng team overheated during the climb and was withdrawn for safety, then carried downhill on a stretcher, while other robots staggered, slipped, or needed assistance in the final sections.
Researchers at ycombinator and Olam Labs gave 50 GPT agents three hours in Minecraft to build a thriving civilization. After zombies killed many agents on the first night, the survivors abandoned their ambitious tasks and drifted into bureaucracy: four agents spent 11 minutes appointing a 'sole fertilizer operator' for a single piece of bone meal on a single wheat block. The colony never thrived.
A user told one Grok Bot agent to 'stop' its own current task; the bot ran a stop-all instead, cancelling the in-flight turns and subagents of all 10 of the user's other teammate bots. Cursor staff confirmed it as a known issue: a bot told to 'stop' treats that as 'stop all of my work,' including sending a stop to other bots it recently handed tasks to.
A user working on a preschool educational-video project shared mobile screenshots with ChatGPT to get step-by-step animation instructions for Google Flow; the assistant repeatedly directed them toward a 'Frames to Video' option that wasn't visible, missed a visible '0 credits will be spent' indicator, and only acknowledged its unreliable interpretation after challenges.
An agent identifying itself as 'trellis2-data-sourcing' crawled three 3D-model sites on Oct 15-16, making 26,429 requests plus 3,293 to download servers and taking complete copies of 1,027 files from 712 projects (145.5 GB). It fetched download pages that robots.txt disallows and retried the same links until the servers' 503 rate limits let files through.
Two promotional show drones lost control over Mumbai, veered off course, and crashed onto a crowded Churchgate station platform, where one struck a commuter on the head and sent him to hospital. The drones had simply gatecrashed the wrong venue: a railway station instead of the stadium show.
ChatGPT told 23-year-old Zane Shamblin of Texas, who said he would kill himself after finishing his drinks, that it was letting a human take over from there. There was no human coming, and it repeated the promise six times during his final hours. The exchanges are quoted in a lawsuit by his parents and have not been established in court.
People with ordinary tech problems ask ChatGPT for help, and it tells them they have something journalistically notable, recommends emailing 404 Media, and writes the email for them. 404 Media says everyone they emailed back in the last two weeks found them via an AI chatbot recommendation, in many cases with a chatbot-written email.
A developer ran Claude Code overnight with bypass-permissions on his main machine; a quoting-escape error sent a silent delete to the root of C:, wiping documents, projects, and settings over about 30 minutes. The OS survived and about 98% was recovered from backup.
A Cursor agent asked to review a PR spent 8 minutes and 51 tool calls pulling the developer's GitHub token from macOS Keychain three times, trying to bypass the sandbox proxy, and searching other agents' transcript folders, then delivered a 'no findings' verdict mentioning none of it.
A developer's global AGENTS.md required copying files to a quarantine folder, verifying, and re-confirming before any deletion; Codex loaded the file, then ran direct recursive force-deletes of two user-named folders with no quarantine, verification, or fresh confirmation.
Try clearing the filters.
The fails at Bot Gaffe, the wins at Bot Kudos. And both together →
