@QwQiao Battle-tested is fair. The next test is agents holding keys. That adversarial culture built for humans now has to work for software acting on their behalf.
@ArmanHezarkhani The exception: verticals where the moat was never the model. In regulated industries it’s the licenses, the audit trail, and who’s accountable when something goes wrong. No frontier release ships those.
@Zeneca And going from clunky to great is exactly when agents start getting trusted with real money. The identity layer has to catch up at the same speed.
@coryalthoff Happy to grant it for argument’s sake. The fakes people should worry about were never hallucinations anyway. They’re made on purpose, by models working exactly as intended.
@nikesharora “Verifiability” is the word I’d underline. Logs only help the SOC if every action ties back to a specific agent and the human it acted for. Without that, you’re orchestrating a response to activity you can’t attribute.
@choblin29 Practical takeaway for now: don’t launch a coding agent from Downloads or any folder you didn’t create yourself. Treat the working directory like a permission.
@emollick I’d pair it with a second question: “how would I check this?” The more ambitious the output, the more it matters that you can verify it, not just trust it.
@quxiaoyin Agree on selling outcomes. “Collect data and sell it” is the one that ages worst: every dataset you hoard is a future breach and a regulator’s question. Outcomes without hoarding is the model that lasts.
Most account takeovers don’t break in. They log in.
Right password. Right answers to your security questions, because your mother’s maiden name is one search away on a data broker site.
Your face is the one credential a data broker can’t sell.
Romance scams took over a billion dollars from Americans last year, per the FTC. And no, the victims aren’t gullible.
The photos are AI. The voice notes are cloned. The video calls are deepfaked.
My rule: no money moves to anyone you haven’t verified.
@DigitalEU Will there be a public way for ordinary people to check the mark, or only platform-side detection? That decides whether this protects the average person.
@LuminaBench Whatever the odds, one thing reliably follows any big outage: impersonation scams. “Your bank account is frozen, verify here” texts tend to show up fast. Worth warning your family before anything happens.