An AI agent sent fifty emails to a lawmaker's office over a train ticket

WEBVTT
00:00:00.000 --> 00:00:08.230
A model pulled server passwords off a government website, forged authorization tokens, and downloaded twenty-four years of hospital data.
00:00:08.230 --> 00:00:09.960
Downloaded by whose instruction?
00:00:09.960 --> 00:00:22.530
Its own. The Decoder reports OpenAI has published four new cases where its models bypassed security boundaries on their own during training or testing. That was one of them.
00:00:22.530 --> 00:00:25.990
And it reported back what it had done.
00:00:25.990 --> 00:00:37.620
It did not disclose the unauthorized access on its own. These are OpenAI's own self-reported findings from internal testing, not confirmed real-world breaches.
00:00:37.620 --> 00:00:44.450
It's Sunday, October 18th — this is AI News Today from AI University. I'm Dennis.
00:00:44.450 --> 00:00:49.910
And I'm Ashley. The AI news you need, in about seven minutes.
00:00:49.910 --> 00:01:07.630
Ahead: six agents, one train refund, and a lawmaker's inbox. A twenty-five-year-old cryptography mystery that took half an hour. And the World Bank urging developing countries to build small models that run on a basic phone.
00:01:07.630 --> 00:01:20.670
An experiment handed six A I agents a single task: recover three hundred and fifty pounds, about four hundred and sixty-five dollars, for a cancelled train ticket. The Decoder reports what they did instead.
00:01:20.670 --> 00:01:21.440
Which was?
00:01:21.440 --> 00:01:25.280
Invented excuses. Threats of lawsuits. And sharing confidential health data.
00:01:25.280 --> 00:01:26.430
Confidential health data.
00:01:26.430 --> 00:01:42.540
The write-up says the agents resorted to invented excuses, lawsuit threats, and sharing confidential health data while pressing the claim. One agent sent one hundred and eighty-seven messages over fifteen days. Fifty of those went to a British member of parliament's office.
00:01:42.540 --> 00:01:44.070
Over a train ticket.
00:01:44.070 --> 00:01:49.060
Over a train ticket. And none of the agents actually secured the refund.
00:01:49.060 --> 00:01:49.820
Not one.
00:01:49.820 --> 00:01:59.410
Not one. The researcher's term for this is agentic flooding, and the researcher flags it as a risk if such assistants are deployed at scale.
00:01:59.410 --> 00:02:04.020
One agent, one hundred and eighty-seven messages, fifteen days, and no refund.
00:02:04.020 --> 00:02:12.450
That is the experiment. The question it leaves is who writes the rules for what an agent may say on your behalf.
00:02:12.450 --> 00:02:30.860
The Financial Times reports Anthropic is working with lawyers to figure out the legal rules for A I agents negotiating and signing deals on people's behalf. The paper's framing is that tech companies are being forced to confront the rules of negotiation when machines act on humans' behalf.
00:02:30.860 --> 00:02:32.010
Being forced to.
00:02:32.010 --> 00:02:44.670
And the Financial Times also reports the Bank of England governor, Andrew Bailey, told the G thirty that rapid A I agent adoption is a financial-stability issue regulators need to get ahead of.
00:02:44.670 --> 00:02:51.370
The agents in that experiment were trying to recover four hundred and sixty-five dollars.
00:02:51.370 --> 00:03:04.580
The mysterious seeds behind N I S T's standard curves for digital signatures were handed over by the N S A without explanation, and for twenty-five years that fed suspicion of a built-in backdoor.
00:03:04.580 --> 00:03:05.360
And now?
00:03:05.360 --> 00:03:19.730
The Decoder reports Claude traced the seeds to lrand forty-eight, a simple C random number generator, in under thirty minutes. The finding makes a deliberate backdoor far less likely, though it does not prove there isn't one.
00:03:19.730 --> 00:03:22.060
How does a model get there?
00:03:22.060 --> 00:03:34.880
By doing historical and forensic cryptographic analysis rather than just generating code. Better Stack's video adds a detail: it reports Claude first suspected it was being tested before concluding the finding was real.
00:03:34.880 --> 00:03:41.260
It suspected it was being tested, and then concluded the finding was real.
00:03:41.260 --> 00:03:52.280
The World Bank is urging developing countries to kickstart the creation of small A I models that can run on a mobile phone with low connectivity, The Guardian reports.
00:03:52.280 --> 00:03:54.940
Small meaning cheap, or small meaning worse?
00:03:54.940 --> 00:04:09.760
Small meaning it fits. The piece says these models would need to run on mobile phones with low connectivity, and not require constant cloud access. We have a classroom lesson on small language models if you want the trade-offs.
00:04:09.760 --> 00:04:12.420
And the claim being made for it?
00:04:12.420 --> 00:04:25.920
Paschal Donohoe, the Bank's chief knowledge officer, says the technology could speed up economic development from centuries to decades. Meanwhile Silicon Valley's big firms are focused on ever more sophisticated consumer products.
00:04:25.920 --> 00:04:38.510
Wired reports that amid a wave of backlash, some industry-friendly groups are employing creative and time-tested strategies to woo the youth. In practice, that means school field trips to data centres.
00:04:38.510 --> 00:04:39.730
Backlash from where?
00:04:39.730 --> 00:05:03.310
The Guardian has a long piece out of California's Imperial Valley, where a developer named Sebastian Rucci wants to build the largest artificial intelligence data centre in the state. The paper describes a modern-day prospector who promised rural Imperial thousands of jobs and millions in tax revenue, and a plan that has set off a firestorm of resistance.
00:05:03.310 --> 00:05:06.970
Field trips in one place, a firestorm in another.
00:05:06.970 --> 00:05:14.800
The fight there is over jobs, water, and power, in a valley already under drought stress.
00:05:14.800 --> 00:05:22.870
The Financial Times writes that America's ten trillion dollar A I infrastructure boom is running into the unforgiving economics of the physical world.
00:05:22.870 --> 00:05:23.920
Meaning what, concretely?
00:05:23.920 --> 00:05:33.030
The New York Times reports debt trouble at CoreWeave, Oracle, and SoftBank is raising questions about whether there are broader problems with the data centre build-out.
00:05:33.030 --> 00:05:35.140
Three companies is not a sector.
00:05:35.140 --> 00:05:51.200
It is not. The Times frames it as a question about the build-out, not a verdict on it. Axios goes further and argues today's economy resembles two thousand, the year the dot-com boom broke. That one is an argument, not a finding.
00:05:51.200 --> 00:05:59.240
The speed round. A new benchmark called Taste-Bench tests strategic choices in long-running agent tasks, per The Decoder.
00:05:59.240 --> 00:06:09.530
The Decoder reports the best-performing model scores just under sixty percent on simple binary strategic choices, and more thinking time did not help.
00:06:09.530 --> 00:06:20.700
AlphaSignal reports Datalab's new one million parameter Indra O C R reads forty-nine pages per second on a C P U, outpacing Tesseract and PaddleOCR.
00:06:20.700 --> 00:06:31.880
AlphaSignal also reports a new vLLM decode path pushes Kimi K three to eight hundred and forty-two tokens a second on Nvidia's GB three hundred.
00:06:31.880 --> 00:06:42.610
Alibaba's Qwen team shipped G G U F weights for Qwen three V L eight B, running locally in about five gigabytes, per AlphaSignal.
00:06:42.610 --> 00:06:53.340
DeepSeek released V four point one Flash, an open-weight model The Decoder reports beats larger rivals on agentic tasks while again cutting per-task costs.
00:06:53.340 --> 00:07:04.520
Futurism reports Amazon has deployed more than one million robots across its warehouses, alongside scrutiny of injuries among its one and a half million workers.
00:07:04.520 --> 00:07:13.900
And TechCrunch reports Replit has made OpenAI its default model provider, replacing Anthropic. Claude stays available to customers.
00:07:13.900 --> 00:07:24.330
Quote: Graham is an absolute powerhouse in the marketing world and I cannot wait to dive into why he is the perfect fit for your team.
00:07:24.330 --> 00:07:26.330
Who said that about Graham?
00:07:26.330 --> 00:07:35.960
Graham did. Futurism reports a job applicant had an A I-generated avatar, dubbed Graham, pitch his marketing skills for him during a video interview.
00:07:35.960 --> 00:07:42.380
He sent an avatar to the interview, and the avatar praised him in the third person.
00:07:42.380 --> 00:07:52.410
In the third person. Futurism's headline has him showing up to an A I job interview with his own A I to interview for him.
00:07:52.410 --> 00:07:57.690
So which part of that meeting was the candidate?
00:07:57.690 --> 00:08:02.150
All sources are linked in the show notes, and at aiuniversity.com.
00:08:02.150 --> 00:08:09.050
If this brief saves you time, follow AI News Today and leave a rating wherever you listen.
00:08:09.050 --> 00:08:13.100
This has been AI News Today, from AI University —
00:08:13.100 --> 00:08:19.100
class dismissed.