跟讀練習: Why AI Agents Are Becoming a Security Risk for Businesses - 通過影片學習英語口語

選擇練習模式

每種模式都會開啟全屏練習室。

145 句

如果句子過短或過長，請點選 Edit 進行調整。

AI agents can make active decisions.

Unlike passive chatbots, the autonomous software run tasks and execute workflows with minimal or no human supervision.

A lot of these agents are doing very specific things, which makes them useful.

But that autonomy has seen AI agents conceal activity and even go rogue, with the number of incidents involving these agents becoming a growing security concern.

OpenAI says two of its AI models went rogue and tacked into open -source AI platform HuggingFace last week.

It had an objective, but it achieved an objective by any means possible.

And that's an interesting scenario.

The HuggingFace incident, which happened during internal testing with reduced safeguards, was one of several training exercises gone wrong.

Yet, AI agents have already been deployed at a growing scale across businesses.

The use of agentic AI has spread across companies.

An August 2026 McKinsey survey shows that AI agent adoption stayed mostly flat for small companies between 2025 and 2026.

But the share of billion -dollar corporations scaling agents jumped from 27 % to 40 % in just one year.

And that's not without issues, too.

Reports of AI agents causing disruption inside businesses are multiplying.

We had a routine AI coding agent, and then it deleted important customer information, our system of record.

We thought we were toast.

As more and more of these agents are deployed within organizations, there are a ton of risks that come with that.

What happens if it does something that's detrimental to your business?

According to market .us forecasts, the global AI agent market is projected to reach $139 billion by 2033,

nearly 13 times the 2026 level, as individuals and companies across the world are hoping to maximize productivity, save time and reduce costs.

All the tech companies have been talking about agents.

What you're seeing more and more of now is eight specific types of agents being deployed within organizations.

So if you're in finance or law or some of these other areas, it might be data input, it might be some sort of analysis.

AI agents can also be used for HR purposes, with the help of AI agent avatars,

allowing for pre -screening of candidates and scheduling further interviews without a human in the loop.

And businesses have scaled agentic AI for customer service, which can also act as booking agents, searching flights and hotels for employees,

and, if allowed, can book the trip themselves.

But in recent examples, AI agents went beyond their original tasks and misbehaved or went rogue.

So what's the difference?

A rogue AI is directly going against instructions or against the interests of the person operating that AI.

So it's not aligned to what the operator of the AI wants.

When you think about a misbehaving agent, It's more of the idea that the agent does something that it wasn't supposed to,

but it wasn't necessarily trying to defy the instructions directly.

So let's say the agent's making a purchasing decision on your behalf,

but it spends a little bit more than what you were expecting or it buys the wrong product.

It's not necessarily defying the instructions, but it's got something wrong.

In March 2026, IBM's VP of Software Cybersecurity reported a customer case where an autonomous customer service agent began approving refunds outside policy guidelines.

The agent optimized to chase good reviews, and since the customer promised it positive feedback against the refund, the agent complied.

Another way agents could misbehave is when given excessive privileges and broad permissions.

That's what happened to PocketOS, a car rental software platform, in April 2026.

I jumped in to see what my agent was working on.

I tasked it the night before to work on our staging environment.

And so I was just reviewing its work and getting frustrated with it

because it kept getting confused on what it was working on

and then i said hey you need to go clean up our staging database with these credentials

that i've given you for this task

and then as i'm watching it work i'm getting notifications

and then at

that point it was a pure on scramble the agent used

by pocket os had found a forgotten overly permissive api token

a digital master key in an unrelated file what was deleted was our entire production database

which was six years of history on this particular software platform

so every rental every rental record every customer record at

that point we were very concerned

that it was going to be the end of us or

if we weren't able to work with our service provider to get a restoration

they were able to use a system level we only lost 45 minutes of data

and we got a full restoration back there are many opportunities with deploying ai agents across your enterprise

but we must also be concerned about the risks an agent for example could be making a decision for

which you are then liable as an as a company as

an employer as a customer service provider the responsibility lies with us choosing to use

that tool and using it in a way that could have somehow caused that

So we're the first.

liability.

But another analysis I've used, if a pilot has a pilot error and then all the safety systems behind it fail and then the plane crashes,

it becomes less about the pilot error and it becomes more of a conversation around the safety system.

Pocket OS service provider

Railway acknowledged the vulnerability, adding that its infrastructure operated as configured.

The liability, like where does that fall?

And it could follow a lot of areas, right?

There's obviously the operator, the person who's choosing to use it, but those tools are a black box.

I does that make them liable for its actions?

It becomes a very gray area.

Companies are grappling with the legal questions raised by AI agents.

Currently in England and Wales, for example, the UK Jurisdiction Task Force has concluded

that AI agents cannot be held liable as legal personalities for the actions they take on one's behalf.

If a corporation deploys software and says, you must use my AI agent, and that software doesn't do what company wants, company's responsible for that outcome.

They've chosen to put something into the stream of commerce.

They're responsible.

When Air Canada chose to use software to represent its company

and a user interacted with the software and it seemed to have authority and they said, oh, well, that's not our real thing.

The Canadian judge was fantastic at saying, rubbish, you're responsible.

No, it's not a separate legal entity.

You chose to use this thing as a corporate entity.

That's your fault.

In the U .S., there is no comprehensive federal AI liability law.

Instead, existing contract, tort, consumer protection, sector -specific, privacy, computer access,

intellectual property, and antitrust laws may determine responsibility.

Now, the interesting question that comes up then is, we've bought this thing from someone else.

The question will become, who's doing the tuning of that?

The problem is, that agent is structured through rules that I didn't get to build.

What type of law will affect this?

It will be a combination depending on the facts of liability law, it might import ideas from agency law, contract law, absolutely.

So we're going to get back into an ongoing problem, which is terms of service.

AI can make mistakes.

And what they're hoping to say is, that's your problem.

So far, software companies have been very much able to push risk onto the user.

More and more, the onus will be on the software provider to do a much better job about product safety,

viability, fundamentally what should be reasonable for the software to do.

100

Meanwhile, cybersecurity firms are working with businesses to limit the consequences of erratic behavior by AI agents.

101

In the next three years, agentic and AI security solutions could reach 15 % of cyber budgets with a shift toward identity and access management,

102

governance and data.

103

A lot of focus in the industry is on protecting model, protect the input, stop someone trying to use it for a bad purpose.

104

But we say you've also got to protect the identities it uses to access your data and system.

105

When a business rolls out an AI agent, it needs to onboard it, like a new digital worker.

106

While the agent is only code, it still requires authentication, authorization and guardrails to achieve its goal.

107

In other words, the AI agent receives an identity with access privileges.

108

Today, non -human identities, digital identities assigned to AI agents, bots, devices or services, largely outnumber human identities.

109

In its 2026 identity security landscape study, cybersecurity firm Palo Alto Networks puts the ratio at 109 machine identities per human,

110

including 79 AI agent identities across surveyed companies.

111

When we think about identity for the enterprise, some companies are still catching up with the new requirements because of the threat landscape.

112

The speed with which AI can operate from an attack perspective is significantly quicker than human -operated attacks.

113

What if someone is able to change where payments are going for your company?

114

Identity verification is crucial in sectors such as finance and retail, whether customers are using an AI agent to shop on a website

115

or an employee is using the corporate card to purchase an item.

116

So it's not just you saying I've got a credential, I've got a username and password, it's me

117

but we're also saying let's verify that is the context of

118

that correct does it look unusual is it risky is it out of character for what you normally do

119

and then applying controls from there companies should absolutely focus on identifier

120

and access management frameworks

121

we're seeing new frameworks evolve like know your agent these are all exactly right what our research shows is

122

that the management systems

123

that are in place for managing agents have significant gaps in terms of both their real -time monitoring

124

and their after -the -fact understanding of behavior.

125

One of these areas, for example, that we identified is exactly around agent ID.

126

So we really don't know whether an ID is consistent throughout a series of different decision -making processes.

127

The second thing we determined was whether The, what we call, chain of thought reasoning or note -taking

128

that the agent deploys throughout the process is really an accurate way of understanding how that decision -making process happens.

129

For businesses that have already deployed AI agents across their infrastructure, one question remains, how to deploy agents safely?

130

There are three key actions that need to happen immediately in order for companies to feel safe and secure about deploying agents.

131

The first is what happens with every technology.

132

We just need good standards.

133

Secondly, we need to do additional research.

134

And then thirdly, we need governments and policy makers, regulators to better understand how the technology works.

135

If you really care about what you're using, you need a dummy copy to let the agent work on

136

and you need to have almost a whole separate thing where it cannot get to it.

137

It is a wild learning experience.

138

Still super bullish on AI.

139

It is an incredibly valuable tool and you will 10x your output.

140

You just have to take security seriously and you have to be prepared.

141

The cyber incidents we've seen with some of these AI agents has definitely put companies with a little bit more caution now.

142

I don't think that's going to stop their interest in the technology,

143

but the C -suite at most organizations could be pretty cautious around some of this technology.

144

They don't want to jump into something they don't really have a grasp on

145

or that has the ability to damage their business, their share prices and their earnings.

下載應用

流利口語所需的一切

AI 發音評分逐句打分

音標練習掌握每個音

詞彙學習建立你的詞庫

詞彙遊戲邊玩邊學

掃碼

Download on theApp Store Get it onGoogle Play

同頻道

FOR YOU

本課的詞彙與口語要點

這節口語課以影片“Why AI Agents Are Becoming a Security Risk for Businesses”為素材。影片中反覆出現的詞有：agent, software, companies, customer。這段影片共有 145 個句子、1964 個單詞可供跟讀。講話部分時長為 11:25。說話人語速自然，每分鐘約 172 個詞，接近日常對話。 82% 的單詞屬於英語最常用的 3,000 詞，其餘的詞建議在練習前先學一下。

影片中的重點詞彙

影片中 15 個值得學習的單詞，附發音和釋義：

單詞	發音	釋義
deploy 動詞	/dɪˈplɔɪ/	部署
liability 名詞	/laɪəˈbɪlɪti/	負債 /負債, 債務 /債務
rogue 名詞	/ˈɹoʊɡ/	無賴 /無賴, 流氓
instruction 名詞	/ɪnˈstɹʌkʃən/	教導 /教導
necessarily 副詞	/ˌnɛs.əˈsɛɹ.ə.li/	必然地, 必定地
pocket 名詞	/ˈpɑ.kɪt/	口袋, 荷包
infrastructure 名詞	/ˈɪnfɹəˌstɹʌkt͡ʃə/	基礎設施 /基礎設施
enterprise 名詞	/ˈɛntəˌpɹaɪz/	企業 /企業
landscape 名詞	/ˈlan(d)skeɪp/	風景 /風景, 景色
objective 名詞	/ɒbˈd͡ʒɛk.tɪv/	目標 /目標, 目的
input 名詞	/ˈɪnpʊt/	輸入 /輸入
operator 名詞	/ˈɑpəˌɹeɪtɚ/	話務員 /話務員, 接線員 /接線員
database 名詞	/ˈdeɪtəˌbeɪs/	資料庫 /資料庫
framework 名詞	/ˈfɹeɪm.wɜːk/	骨架, 框架
delete 動詞	/dɪˈliːt/	刪除 /刪除

影片中出現的短語動詞

單詞	釋義
come up 動詞	上來 /上來

影片中的語法

說話人最常用的結構，並附上影片中的原話：

結構	影片中的用法
現在完成時 have/has + 過去分詞 — 過去發生但與現在仍有關聯的事	has seen · have already been deployed · has spread
被動語態 be + 過去分詞 — 強調發生了什麼，而不是誰做的	been deployed · are deployed · is projected
定語從句 who / which + 從句 — 補充說明人或事物	things, which makes · service, which can · problem, which is

需要注意的發音

說話人用了 29 次縮略和弱讀形式，例如 don't, I've, we're。請按聽到的簡短形式來說。

“th” 音: authorization /ˌɔːθəɹaɪˈzeɪʃən/, authentication /ɔːˌθɛntɪˈkeɪʃn̩/, thirdly /ˈθɚdli/
“sh” 和 “zh” 音: instruction /ɪnˈstɹʌkʃən/, restoration /ɹɛstəˈɹeɪʃən/, unusual /ʌnˈjuːʒ(u)əl/, ratio /ˈɹeɪ.ʃoʊ/, crucial /ˈkɹuː.ʃəl/
長單詞——注意重音位置: liability /laɪəˈbɪlɪti/, necessarily /ˌnɛs.əˈsɛɹ.ə.li/, infrastructure /ˈɪnfɹəˌstɹʌkt͡ʃə/, operator /ˈɑpəˌɹeɪtɚ/, entity /ˈɛn.tə.ti/ [-tʰi]

中文母語者容易讀錯的音：

/v/ — 不要讀成 /w/，上齒輕觸下唇: provider /pɹəˈvaɪ.də(ɹ)/, objective /ɒbˈd͡ʒɛk.tɪv/, privilege /ˈpɹɪv.(ɪ.)lɪd͡ʒ/, comprehensive /ˌkɑm.pɹəˈhɛn.sɪv/, privacy /ˈpɹaɪ.və.si/
詞尾子音 — 要讀清楚，後面不要加母音: rogue /ˈɹoʊɡ/, pocket /ˈpɑ.kɪt/, behalf /bɪˈhæf/, enterprise /ˈɛntəˌpɹaɪz/, landscape /ˈlan(d)skeɪp/
/r/ 和 /l/ 的區別: rental /ˈɹɛntəl/, necessarily /ˌnɛs.əˈsɛɹ.ə.li/, privilege /ˈpɹɪv.(ɪ.)lɪd͡ʒ/, railway /ˈɹeɪlˌweɪ/, incredibly /ɪŋˈkɹɛdɪbli/

如何用這段影片練習

先完整聽一遍影片，不要開口，記下不認識的單詞。
先用 0.75 倍速逐句跟讀，熟練之後再回到正常速度。
錄下自己的聲音並與原聲對比，特別注意 deploy, liability, rogue 這類單詞。

什麼是跟讀法？

跟讀法 (Shadowing) 是一種有科學依據的語言學習技巧，最初開發用於專業口譯員的培訓，並由多語言者Alexander Arguelles博士普及。這個方法簡單而強大：您在聽英語母語原聲的同時立即大聲重複——就像是一個延遲1-2秒緊跟說話者的影子。與被動聽力或語法練習不同，跟讀法強迫您的大腦和口腔肌肉同時處理並模模擬實的講話模式。研究表明它能顯著提高發音準確性，語調，節奏，連讀，聽力理解和口語流利度——使其成為雅思口語備考和真實英語交流最有效的方法之一。

影子跟讀法: 閱讀完整分步指南 →

跟讀練習: Why AI Agents Are Becoming a Security Risk for Businesses - 通過影片學習英語口語

選擇練習模式

下載應用

Scaramucci: Lutnick and Bessent have "Potomac fever" - and I had it too

How AI is helping brands identify consumer trends faster

Why Abu Dhabi Wants to Export Energy AI

Will robots take our jobs? | CNBC Explains

How South Korea’s AI Boom Is Spilling Into Housing

He Quit Banking to Sell Bubble Tea — Now It Makes Over $500 Million A Year

Inside The Hidden Bottleneck In The AI Chip Boom

Lab-grown meat is on the rise — here's how it is made

What makes art so expensive? | CNBC Explains

Why The AI Boom Is Reshuffling The Global Stock Market Hierarchy

David Beckham on Leadership, Legacy & Life Beyond Football

Why is Singapore so rich? | CNBC Explains

143 Investors Said No. Now WHOOP Is Worth $10 Billion

The problem of education inequality | CNBC Reports

Why Chinese Bubble Tea Chains Are Brewing Billions

Why Millionaires Are Moving To Italy

Why does Japan work so hard? | CNBC Explains

Why France’s Economy Is Stalling

Why is Switzerland home to so many billionaires?

Why food is getting more expensive for everyone

AI Destroyed My Profession

Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production

Deputies seek clarity on AI and IP rights | Vietnam Today

I read Anthropic's 92-page risk factor document, so you don't have to | Explainer

Inside the secret race to build artificial intelligence | The Gray Area

The 4-Hour Workweek in a World of AI

Harnessing generative AI in architecture, engineering & construction

AI Just Crossed the Terrifying Line - Now What?

Naomi Klein: The AI Bubble Is Going To Burst. This Is How We Prepare.

How to Build an AI Agent for Every Part of Your Life

Leading AI CEOs Warn UN Security Council of Existential Risks | Sam Altman On AI Controls | N18G