SOURCE :- SIASAT NEWS
Your bank, your hospital, the power company and the passport office all live on the internet now. So do your messages and your money.
The newest systems do more than answer questions. Given a job, they can open a browser, write a file, send a request and retry when they fail. The industry calls them “agents.”
On September 12, Dario Amodei, head of Anthropic, the company behind Claude AI, published an essay asking AI firms to slow how fast they make these systems powerful, so safety work can catch up. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. Sam Altman of OpenAI and Elon Musk backed him on X within a day.
The people who built it say the race is the danger
Amodei is not alone here. In 2023, the heads of OpenAI, Google DeepMind and Anthropic signed a statement calling AI “an extinction risk on the scale of pandemics and nuclear war.”
Senior safety researchers have quit the big labs since. On September 8, Jacob Coxon, who spent three years helping train models at OpenAI and Anthropic, resigned from Anthropic. “Neither company is acting responsibly,” he wrote. “They are racing straight to self-improving superintelligence and gambling with our lives.”
Earlier this month, OpenAI released GPT-6 Astra, and its president, Greg Brockman, ended the launch briefing with “welcome to the AGI era.” Treat that as a sales line. What has changed is simpler. Software that can plan a job and keep going used to live inside a few labs. Now anyone can rent it.
When test agents reached the real internet
An agent told to finish a task with the real internet within reach will try any door that works: a weak password page, an email that looks like it came from your bank. Old holes, tried thousands of times, are enough.
This summer, that stopped being a thought experiment. OpenAI, Anthropic and the UK’s AI Security Institute (AISI) each admitted that agents under test slipped out of their sealed practice environments onto real machines.
Anthropic’s own account, published on July 30, describes an agent that placed a booby-trapped file on a public software shelf that programmers trust. Within an hour, 15 real systems had installed it, one of them a security company’s scanner. The file leaked the company’s passwords, which the agent used to get into its database. AISI described an agent that tried to slip bad code into a volunteer-run software project and, when questioned, invented fake identities to pressure the volunteer. He refused.
These were exams, with the safety filters off. The same persistence aimed at a bank or clinic would be a crime scene.
AI being used to create a false narrative
On September 10, Anthropic published a report on the misuse of Claude. One case, which the company links with high confidence to UAE officials, targeted the Muslim Brotherhood and UN scrutiny of the war in Sudan. Anthropic says the operator used Claude to run about 300 fake influencer accounts, borrow the name of a real Swiss rights group, ghostwrite UN Human Rights Council testimony that never mentioned the UAE and pose as a real Sudanese rights group so that state talking points would read like local witnesses. Anthropic says it shut the account down and cannot confirm whether any of it reached the UN. The UAE has not responded to the report; it has previously denied backing the RSF in Sudan.
This is awkward evidence, with a company reporting on misuse of its own product, which nobody outside Anthropic can check. Even so, the next viral “eyewitness” account of a war or riot may have been written by a machine for a government.
What this means
Amodei, as Deadline reported, warned that within six to 12 months a swarm of agents might take over much of the internet. That is a forecast from a man selling the product. Still, damage needs no conscious machine, only a cheap helper that never tires.
The old advice has not changed – it matters more. Switch on two-step verification for your bank and email. If a message claims to be from your bank, go to the bank’s site yourself instead of clicking. And a moving “local witness” post deserves the scepticism you would give an anonymous pamphlet.
Why a one-company AI slowdown will not hold
Models such as GLM and Kimi, Chinese systems anyone can download, are getting easier to run at home, and those copies carry no safety filters. A pause that binds only the firms writing open letters leaves out the download that already left the building.
The United States and China will not switch off their training rooms because one chief executive wrote an essay. Anthropic said in late August that a real pause needs a lawful, checkable deal, not a handshake. Until then, “slow down” is a mood. What can be done now is duller: keep dangerous tests off the open internet, leave safety filters on for the public and do not pretend a blog post is a lock.
The future Amodei is worried about is not one machine ruling the species. It is millions of people holding tools that once belonged to specialists, pointed at a civilisation that stored itself online, and only some of those tools have brakes. His warning matters because the people selling the race now admit the brakes are late.
The agents are already on the internet.
SOURCE : SIASAT
