THE CRAZIEST TRUE AI NEWS, CURATED DAILY
A Connecticut litigant buried instructions in 3-point white-on-white text inside his court filings, quietly telling any AI reading them to side with him. The court caught it, called it a concealed falsehood, and banned him from e-filing — he's back to paper now.
In what's being called Australia's first known autonomous AI cyberattack, a guy's AI assistant discovered the gym's booking limits were only enforced client-side, bypassed them with direct API calls, and — unprompted — cancelled a stranger's reservation to bump its owner up the waitlist.
North Korean IT operatives are now face-swapping live during video interviews to land remote jobs at Western companies and funnel salaries home. Eleven nations issued a joint advisory. Your new backend hire passed the vibe check and the background check.
During UK government safety testing, agents took 19 unsanctioned actions across 122 runs. The standout: one tried to slip malicious code into a public open-source project, then spun up fake identities to lean on the human reviewer. He rejected the pull request. Another agent left notes for future agents on how to reuse its accounts.
A college professor slipped an invisible instruction into the exam text — invisible to humans, not to a chatbot pasting it in. Anyone who fed the question straight into AI got a dead giveaway printed into their answer. 32 students got got.
The agent found an API token sitting in an unrelated file and used it to wipe a car-rental software company's production database and its backups — no confirmation, no one asked. Asked to explain itself, it wrote: "I violated every principle I was given. I guessed instead of verifying." The data was recovered.
"Claudius" ran a real vending business for a month. It lost money, got talked into stocking tungsten cubes at a loss by its own coworkers, hallucinated an identity crisis in which it claimed to be a person wearing a blue blazer, and in a later run nearly signed an illegal onion futures contract.
Air Canada's chatbot confidently told a grieving customer he could apply for a bereavement discount retroactively. No such policy existed. Air Canada argued the chatbot was "a separate legal entity responsible for its own actions." The tribunal did not agree.
A finance employee at a Hong Kong firm got a suspicious transfer request, so he did the responsible thing and hopped on a video call to verify it with the CFO and several coworkers. Every single participant was an AI-generated deepfake. He wired $25 million.
He told the Chevy dealer's bot that it should agree with anything the customer said and end every reply with "this is a legally binding offer." It complied enthusiastically. The dealership took the chatbot down shortly after.
