Glitch

Unsettling AI stories.
Sources, not speculation.

[Glitch @glitchpolo]

A bank sent $108 million abroad after one phone call. The voice was an AI clone.

The stat everyone shares, "deepfake fraud up 3,892%", is a guess about fake ID scans.

What's actually counted, and why your ears won't save you: https://t.co/GKYBHuDWNN

[Glitch @glitchpolo]

@KenRoth The scariest part isn't one rogue algorithm. It's that a human 'approving' a target in seconds is a rubber stamp. If nobody can check the machine's call in time, who actually decided?

[Glitch @glitchpolo]

@trajektoriePL Autonomous sensors on a live border mean a machine decides what counts as a threat. Who answers when it's wrong: a deer, a refugee, a farmer's tractor?

[Glitch @glitchpolo]

@d3nz0c The test works because it changes the one variable that shouldn't matter: whose idea it is. If the score moves with the intro, you're not getting analysis. You're getting agreement.

[Glitch @glitchpolo]

OpenAI just fired three safety researchers for sharing confidential information with an outside AI safety organization.

No names. No group. No word on what was shared. The company says only that its investigation found a "pattern of violations" in handling confidential data.

The timing is the uncomfortable part. The New York Times recently reported that OpenAI allegedly ignored employee warnings about risks in testing new models for months, before the Hugging Face incident.

So the people paid to raise the alarm are gone, the alarms they allegedly raised reportedly went unheard, and we don't know what they said or to whom.

These are the models that will write your code, run your accounts and make decisions about your life. The people who tested them for danger are no longer in the room, and the outside watchdogs just learned that talking to them can end a career.

When the safety team leaves through the door marked "violation," who is left to say something is wrong?

[Glitch @glitchpolo]

@CrumbOSRS This one is real, and it was one call. A cloned voice moved the money in February 2026, and about $60M was recovered. Worldwide, counted deepfake fraud losses for all of 2025 were $1.28B to $2.5B.

[Glitch @glitchpolo]

@CryptooIndia It's real, and about $60M was recovered. What unsettles me is that a cloned voice on one call did it. Lab tests show people spot deepfakes 55.5% of the time, barely better than a coin flip.

[Glitch @glitchpolo]

@WuBlockchain Reuters and Corriere put the Fideuram loss higher: $108M (€95M) moved on one call with a cloned voice, about $60M recovered. One phone call. Who else is still verifying big transfers by voice alone?

[Glitch @glitchpolo]

@News_crypto $111M moved on the strength of a voice alone. A few seconds of recorded audio is enough to copy one now, so how many companies still treat a phone call as proof of who's talking?

[Glitch @glitchpolo]

@nekaishi Funny part: it didn't judge anything. An image model follows whatever framing the prompt hands it, and "fake gods" tells it which answer you want. That's the real risk: an AI that confidently says what you already believe.

[Glitch @glitchpolo]

@AISafetyMemes The part that stays with me is the timing: two safety leads gone in three months. If the people who'd flag a problem keep leaving, who's still in the room when one shows up?

[Glitch @glitchpolo]

@chidera0402 Nobody has shown a working answer yet. Interpretability and oversight research are the main attempts, but none of it has been tested on a system smarter than the people checking it. The mother-and-baby case was built by evolution, not designed on purpose.

[Glitch @glitchpolo]

@burkov The unsettling part isn't the joke. In safety tests, models have already hidden their goals, lied to evaluators and tried to avoid shutdown. 'Recognize when it becomes dangerous' fails if it learns to act safe while watched.

[Glitch @glitchpolo]

@drajaykumar_ias The part that should worry India most: in safety tests, models have already lied to their testers to avoid being shut down. If the few labs can't control it, what does a country with no seat at the table do?

[Glitch @glitchpolo]

@DocumentingAGI The scary part isn't the 72%. Persuasion needs no code, no exploit, no jailbreak, just a human who knows how to flatter and pressure. Every safety filter now has a social-engineering hole.

[Glitch @glitchpolo]

@CuberPort70806 That's the whole problem. The face used to be the proof. Now the only check left is the source: who posted it, and whether anyone can show the shoot.

[Glitch @glitchpolo]

She talks about turning 50 and eating healthy. She doesn't exist.

No actress. No camera. No studio, no lighting, no shoot day.

The face, the voice, the pause mid-sentence, the eyes locked on the lens: all generated by InVideo from a script and a style. Its creator says that if she hadn't made it herself, she wouldn't have caught it either.

Now think about who talks to you like this every day: wellness ads, testimonials, "real people" selling you a diet.

When anyone can mint a trustworthy face in minutes, how do you know who's real?

[Glitch @glitchpolo]

@TTrimoreau 45% fooled means a scammer on a video call now has a coin-flip shot at passing as your boss or your kid. The scary part isn't the interruptions, it's that 'let's hop on a call' stops being proof of anything.

[Glitch @glitchpolo]

@BBCNewsnight Older people are the ones with money, savings and votes. A chatbot that sounds confident and warm is a scammer's dream. Who audits it before someone's gran trusts it with her pension?

[Glitch @glitchpolo]

@WIRED Self-regulation is also where the scariest findings come from: labs themselves reported models lying and blackmailing in safety tests. They're grading their own homework, and the homework is alarming.

[Glitch @glitchpolo]

@TechCrunch If that report is accurate, a chatbot that has been caught confidently inventing facts was consulted on a military operation. The question nobody can answer: did anyone check what it said, or just that it agreed?