Humans Meet Tech · Interview
AI Made Cyberattacks Smarter. Are We Ready to Recover?
A conversation with Yasmine Kamara, cybersecurity and IT risk expert, on why the AI that makes phishing emails nearly perfect has made recovery, not prevention, the real test of an organization’s security.
Cybersecurity · Governance · Responsible AI

Listen to the conversation
Humans Meet Tech · Full episode · 37 min listen
A few years ago, at a multinational company with operations across several continents, someone in accounts payable got an email that looked like it came from a vendor, asking for updated banking information and approval to process an invoice. The employee did not check the usual form, did not route it through the approval chain, and did not confirm the address against the one on file. They just processed it.
The company wired €300,000 to an account that belonged to no vendor at all.
Can you imagine being that person that did that? You’re losing your job and your company might even sue you for that.
That is Yasmine Kamara, describing a case from her time at EY, where she advised clients on technology risk. For years, cybersecurity was framed as a wall, the right tools, strong passwords, enough vigilance to keep the bad actor out. The email that cost that company €300,000 did not look like an attack. It was polished, well timed, and indistinguishable from a routine vendor update, and Kamara says that’s by design. The tools available to attackers have caught up with the tools available to defenders. So the question worth asking now is whether an organization can detect, contain, and recover once an attack gets through.
Kamara has spent more than a decade in technology risk, IT audit, and information security, including roles at EY and SiriusXM, working with organizations on how to keep data safe and how to recover once something goes wrong.
Prevention Was Never the Whole Plan
Ask Kamara what cyber resilience means and she starts with a mindset shift, one she thinks most organizations haven’t fully made.
It’s no longer a matter of if I get attacked, it’s a matter of when I’m gonna get attacked.
For a long time, cybersecurity stopped at prevention. Kamara’s framing adds the half most organizations still treat as an afterthought, a plan for the day prevention fails. That plan lets an organization detect what happened, contain it, and keep operating while it recovers.
The signals employees used to rely on are disappearing too.
You could literally go on ChatGPT and draft a perfect email, have the right tone, and it’s harder to determine that this is coming from a bad actor.
That changes how the wire transfer story reads. The AP employee wasn’t unusually careless. The usual tell, the sense that something about an email feels off, never showed up. So the only thing standing between a normal Tuesday and a six-figure loss is whether the process gets followed, however convincing the request looks.
The Layer Most Companies Build Last
When Kamara breaks resilience into layers, the order surprises people who assume awareness training is where cybersecurity starts. In her framing, it is closer to where it ends.
The first layer is a standardized process with written steps for the mundane moments. What to do when an email looks off, who to escalate it to, and what a legitimate request should look like.
Process comes first because it removes the need for an employee to make a fast, high-stakes judgment call mid-workday. This is exactly the situation the AP employee was in when a five-minute pause could have caught what a perfect-sounding email could not.
The second layer is detection and containment, which means catching what gets through before it spreads.
“Security in depth” is Kamara’s term for stacking these layers rather than relying on any single one. She compares the training piece to fire drills. A plan nobody has practiced does not do much on the day it is needed. Cybersecurity training needs the same realism, with scenarios drawn from how attacks actually happen rather than exercises that are quick and easy to run.
Most defenses still count on every employee spotting every fake. Good security makes the dangerous action hard to complete, even when the fake succeeds.
Resilience Runs on People Who Speak Up
The part of Kamara’s answer that reframed the conversation was about what happens after someone makes a mistake.
Many organizations have built a reporting relationship with their own security team that discourages the exact behavior resilience depends on. An employee who clicks a bad link and is afraid to say so is a bigger risk than one who reports it, because nobody else knows there’s a problem to contain.
The better you let people know, “this is what I did,” the easier, the faster they’ll be able to recover and contain if there’s an actual impact.
Building that kind of reporting relationship means the security team stops looking like, in her words, the police, the people who show up after a mistake to assign blame. Saying something went wrong has to feel easier and safer than staying quiet and hoping it was nothing.
Once employees understand that logic, the responsibility for security stops sitting only with IT.
When people understand that I need to do my part for my colleagues to also be able to do their job, then we know it’s a collective effort.
In Kamara’s experience, this usually breaks down at the top, because leadership sets the tone. When leaders visibly take information security seriously, it becomes easier for employees to follow suit. When they treat it as someone else’s department, the culture underneath the tools erodes no matter how good the tools are.
The Employees You Never Hired
The gap Kamara worries about most is shadow IT, the tools and workarounds employees adopt because the sanctioned option is slower or more annoying. “It makes my life easier, it makes my work faster, but then you’re opening your company to potential risks that they may not even be aware of.”
What used to be called shadow IT is becoming shadow AI, and the exposure is invisible in exactly the same way. An employee uploads a confidential document to a public AI tool because it is faster than the internal system, and there is no reliable way for the company to know it happened.
There’s no way you can tell that someone went and used a publicly available AI platform and uploaded confidential data about a company.
Kamara’s clearest advice for spotting social engineering is to watch for a feeling, the pressure to act fast.
Every time you receive a request like that, you should notice that there’s always some type of urgency. Even if there is, it should go through the normal process.
She has been on the receiving end of this herself. While at SiriusXM, she got a text message from someone posing as the company’s CEO.
I received a text message from someone pretending to be our CEO. If she wants to talk to me, she would email me.
The request that followed was for gift cards, a detail she still finds almost funny in hindsight, because the format of the ask was the giveaway long before the content was. A real executive does not usually text from an unfamiliar number asking for something that converts to cash and disappears instantly.
The harder version is coming. Kamara points to deepfake audio and video as the next escalation, technology that can put a convincing likeness of a real executive on a live video call and ask for a wire transfer in real time. Her answer stays the same for every version of this threat. The signal is the urgency, however real the voice or face seems. Pause, verify through a separate channel, and let the process catch what the moment cannot.
The old tell
- Misspellings and odd phrasing
- A domain that looks slightly wrong
- A tone that feels off
- Depends on the message sounding fake
The tell that still works
- An unexplained sense of urgency
- A request that skips the normal channel
- A demand for secrecy or speed
- Depends on the process, not the message
Keeping Humans in the Loop as AI Takes More of the Decision
The newest version of this question, for Kamara, is agentic AI, systems given access to a company’s tools and told to complete tasks on their own. She sees real, useful cases already, agents that move through hundreds of documents and summarize them before a person makes the final call, saving time that used to disappear into manual review.
Where she draws the line is the decision itself.
It would be very immature and naive of us to just let those agentic AI do whatever they want, ’cause it could be a disaster.
Configuration and access controls matter, and they need governance behind them. The same discipline that makes a wire-transfer process resistant to a convincing fake email is what has to govern an AI agent with access to a company’s systems. The AI should have a defined scope, and a review point before anything irreversible happens, and above all, a person accountable for the outcome.
When technology moves faster than an organization can govern it, the risk moves somewhere harder to see.
A Personal Version of the Same Discipline
Kamara applies the same discipline to her own life.
Do not upload any personal identifiable information, your Social Security number, your ID number, your driver’s license number. Any type of number you wouldn’t feel comfortable sharing with anybody, stay away from uploading it.
Password reuse gets the same warning. Once a hacker guesses a password on one account, the next move is to try it everywhere else, email, banking, social media, until something opens. “Try your best not to reuse passwords. That’s really the biggest area where most people tend to make the mistake.”
Her own habits reflect how far the threat has moved. She no longer answers calls from unfamiliar numbers.
Whenever I get spam calls or any random phone number calling me, I don’t answer, ’cause I’m afraid someone might steal my voice.
Voice cloning is precise enough now to feed a future deepfake or a bank’s voice authentication, which turns answering the phone into a data point worth protecting.
None of this stops a sophisticated attacker outright. The goal is to limit how much damage any single mistake can cause, the same logic that runs through everything else she said.
Resilience Has a Physical Address
It is easy to think of a cyberattack as something that happens only on a screen, but Kamara has visited enough data centers to know the risk has a physical address too.
Getting into one takes identity verification and biometric scanners, and in facilities hosting multiple companies, a locked cage around each client’s servers so no one reaches a neighbor’s equipment. Raised floors keep water away from the electronics during a flood. Fire suppression runs on gas rather than water, so a fire can be stopped without destroying the servers it threatens. Batteries and generators keep power running through an outage.
None of that makes a data center invulnerable, which is exactly why redundancy matters as much as any single safeguard.
They think about redundancy, making sure that your data can be replicated to another server somewhere else in a different geographical location.
That is recovery, built into the infrastructure before anything goes wrong. A resilient organization assumes any one location could fail and builds the ability to keep operating anyway.
A short list for the next budget cycle
- 1
Check the process behind the tools. Is there a written, followable process for wire changes, vendor updates, and urgent executive requests?
- 2
Test the reporting culture. Would an employee who made a mistake report it within the hour, or would they sit on it?
- 3
Extend oversight to vendors. Is anyone accountable for what the company’s third parties do with its data?
- 4
Keep a human on the irreversible step. Where AI agents already touch company systems, does a person still sign off before anything that cannot be undone happens?
The company that lost €300,000 already had a process for vendor changes. A well-written email was enough to get someone to skip it. As AI gets better at sounding like the people we trust, the question worth asking is what happens in the first hour after someone on your team realizes they’ve made a mistake. Do they tell anyone?
Cybersecurity is a task for everybody. It’s not just for your IT team or your information security team. We all play a part in it.
About the guest
Yasmine Kamara
Cybersecurity & IT Risk
Has spent more than a decade in technology risk, IT audit, and information security, with experience at organizations including EY and SiriusXM, advising businesses on how to protect their data and recover when something goes wrong.
This conversation is drawn from the Humans Meet Tech podcast, hosted by Mariam Ammar. Quotations have been lightly trimmed for length and clarity.
Join the conversation
If someone on your team clicked a convincing phishing link today, would they report it within the hour?
Loading comments…
When the Voice on the Phone Isn’t Real