Episode 52 of The Security Podcast of Silicon Valley aired on August 15, 2024, with Jon McLachlan and Sasha Sinkevich hosting. The guest was Suha Can, then the chief information security officer of Grammarly, the AI writing assistant (Podbean, LinkedIn). The episode is called “Suha Can, CISO at Grammarly: Safeguarding User Data in Enterprise AI Systems.” It has no video version.
“the mindset has to be that there must be no mystery, no unanswered questions.”
Suha Can, at 34:50
Suha held two jobs at Grammarly from September 2022 to February 2025. As CISO he led security, privacy, responsible AI, governance, risk and compliance, fraud and abuse, and IT. As vice president of engineering he led the teams that built Grammarly Business for enterprise customers. Before that he founded the Payments Security team at Amazon and grew it from 1 to more than 100 engineers, and he spent more than ten years at the Microsoft Security Response Center. Since March 2025 he has been vice president and chief security officer at DoorDash (LinkedIn).
The tape dates itself to 2024. Sasha says Grammarly was doing AI “before AI was as popular as it is in 2024,” and Suha says Grammarly is launching a customer-held key for its enterprise customers at the end of the month. Grammarly announced that feature, which it calls Bring Your Own Key, on June 25, 2024, for “the coming weeks” (Business Wire, Help Net Security).
This page follows the conversation in order: a section for each part of it with the sources beside it, sixteen figures, twenty-five clips from the tape, the full transcript, and five tools: a quiz, a scorecard for an AI product that wants enterprise buyers, a calculator for the race between a patch and an exploit, a timeline, and a to-do list. Where the decodes of the tape disagree on a word, the transcript puts its best guess in brackets, and the page doesn’t quote those words.
A coffee shop with one computer
Jon asks how Suha got into security. Suha grew up in Turkey in the middle of the 1990s, and the one thing he wanted and didn’t have at home was the internet. He was in high school at the time.
“So I found this coffee shop in a nicer part of town.”
Suha Can, at 2:08
It was far from home, so he took a long walk there after school. On the bottom floor sat the one computer, and through it he found the IRC, Internet Relay Chat, a text chat system built around group channels (Wikipedia). The people on it were mostly students from the few technical universities, plus managers and admins.
“And really, when you look at the time, it was amazing because I was seeking knowledge.”
Suha Can, at 2:52
Everyone else there knew more than he did, he says. Two communities met him there. One ran Unix servers for companies and universities and cared about securing them. The other was the demoscene, the computer art scene that makes small programs with graphics and music to show off programming skill (Wikipedia). Suha joined a Unix security group.
“There was a sense of curiosity, seeking to understand how things work. It was all about having fun, gaining knowledge, a little bit of friendly competition for sure.”
Suha Can, at 3:28
He says several of the people he met there went on to become security founders. His LinkedIn lists him as staff at GSULinuX, on Linux and network security, from 2000 to 2005, and as an enterprise IT security intern at IBM in 2003 (LinkedIn).
Two hats
Jon calls himself a paid Grammarly member of many years and asks what the CISO does there. Suha says he wears two hats.
“Job number one is the CISO job. So I am responsible for all aspects of trust in the company.”
Suha Can, at 4:41
Under the first hat he lists product security, corporate security, privacy, compliance, corporate engineering and the IT infrastructure, and the responsible AI research team. The second hat is engineering. He leads the engineering teams that build the enterprise version, which he says has a lot of interesting security and IT management problems to solve.
Jon suggests that makes him chief product security officer too. Suha agrees and adds chief compliance officer and chief privacy officer.
Solutions that scale
Sasha asks how a security organization with that many parts runs without slowing down research and development. Suha starts with scale.
“So I think one of the important things is really building solutions that scale.”
Suha Can, at 5:59
The first step is to know the company’s risk tolerance. At Grammarly, he says, trust has to be paramount, so the security bar is high, and the budget has to match it.
“Obviously, if you work in a company and security is very important, but you don’t have the right investment in the area and you don’t have the right strategic voice to influence the company, then no matter who you hire, you’re just going to be always behind.”
Suha Can, at 6:25
“And the business moves at a very fast speed. You definitely cannot be in the business of slowing things down.”
Suha Can, at 6:39
So the order is strategy first: a clear vision, the board along, leadership around. Then hiring, then teams. Then a lot of automation, which gets the basics in check.
Necessary but not sufficient
Once the basics are in check, scale stops being enough.
“and I truly believe that scalable solutions are necessary but not sufficient.”
Suha Can, at 7:16
“I always believe that for really high-risk areas or your crown jewels, you need to go a lot deeper. You need to build layers in the defenses.”
Suha Can, at 7:22
That takes skilled people who do manual work on the riskiest systems. He sums it up as two investments in a row: things that scale first, then depth in the areas that matter most to the business.
I didn’t know security was this
The third part is how the team works.
“And I think security has to have a builder mindset. So my teams don’t operate in a way where we advise teams what to do and then we move away.”
Suha Can, at 7:57
Most of his security engineers are also software engineers, he says, and that’s how the team earns credibility. He tells a story from his time at Amazon that, he says, has repeated itself since. His team does a security deep dive on a service and presents its findings to the team that owns it.
“And the principal engineer of the team told me, hey, I didn’t know security was this.”
Suha Can, at 8:38
“But typically, that’s not how you’re going to be able to land impact, and that’s not how you’re going to gain the respect of the engineers, which ultimately are your closest partner in security.”
Suha Can, at 8:58
A builder profession
Sasha agrees that security in isolation doesn’t work and has to be driven by the business, with engineering teams implementing the controls. Suha goes one step further.
“I even go farther than that, in the sense that we actually implement the controls.”
Suha Can, at 9:37
His reason is reuse. If you find a threat and want a control for it, he says, you probably want to build it, because the same control is likely to help other teams in the company.
“We have to have a good mix of security skills and software engineering skills.”
Suha Can, at 9:47
His LinkedIn says the same about his time at Amazon. His engineering organization built and ran the core payments infrastructure that processed all of Amazon’s payments, and his security organization owned automation, application security, red team, threat intelligence, detections and incident response for the payments businesses (LinkedIn).
AI before it was popular
Sasha asks what Grammarly does better than anyone else in the world. Suha’s answer is AI for communication.
“So I would say that Grammarly is the only AI company really focused on transforming how people communicate, whether it’s at work, at school, for personal communications.”
Suha Can, at 10:31
He says Grammarly has been at it for about 15 years, and he puts it at something like 95% of the Fortune 500. Wikipedia dates Grammarly’s launch to 2009, by founders Alex Shevchenko, Max Lytvyn and Dmytro Lider, and its generative AI product to April 2023 (Wikipedia).
“So I think it’s important to highlight that Grammarly has been doing AI before AI was as popular as it is in 2024.”
Sasha Sinkevich, at 11:03
Sasha thanks him for a security team that protects the data everyone gives Grammarly, and says it was just a thank you.
From consumers to enterprises
Jon brings up the shift from consumers to enterprises. He likes the tone feature that shows small emojis for how an email might sound, and then asks the enterprise question.
“But at the end, like, in terms of security, if you sell that product to the enterprise space, usually there’s, like, certification questions that come up.”
Jon McLachlan, at 12:28
Suha draws the line between the two markets. Consumers ask how their data is handled, which is privacy. Enterprises make security a deep topic of its own.
“There’s a lot of questions about how the data is handled.”
Suha Can, at 13:07
One of the ways we communicate
Suha describes certifications as one of the ways companies tell each other that they meet certain standards, with a third party agreeing. He says Grammarly holds the common security certifications.
“We are also quite aggressive in pursuing new certifications, especially in the realm of AI and privacy, because especially with AI, there is a lot of nervousness in the industry.”
Suha Can, at 13:38
The nervousness depends on where you stand. In Silicon Valley, he says, everyone is embracing AI and building on it. The global CISOs he talks to are more cautious.
The next certification Grammarly is going after is ISO 42001, and he thinks Grammarly is one of the first to do it. Grammarly’s compliance page today lists SOC 2 Type 2, SOC 3, ISO/IEC 27001, 27017, 27018, 27701 and 42001 (Grammarly). A SOC 2 report is an audit report under the AICPA’s Trust Services Criteria (Wikipedia), and ISO/IEC 27001 sets the requirements for an information security management system (Wikipedia).
ISO 42001
Compliance matters to a security buyer, Suha says, but it isn’t everything. Later in the conversation Jon and Sasha come back to the standard. Jon says he doesn’t know any company that’s certified yet.
“Yeah, I don’t know any company yet that’s 42001 certified, but I saw that come out a couple of months ago, and I was like, this is great. This is interesting.”
Jon McLachlan, at 17:35
“And it’s important that companies consider the certifications as standards, and frameworks are not there for the check mark purposes.”
Sasha Sinkevich, at 18:09
Grammarly got there in April 2025. On April 15, 2025, it announced its ISO/IEC 42001:2023 certification, which it called the world’s first AI management system standard, and named the certification body as Mastermind (Business Wire). That was about a month after Suha moved to DoorDash.
A dedicated responsible AI team
The other half of the answer is the work that a certificate doesn’t show. Suha starts with the people.
“Responsible AI is an area where we have a dedicated team, a robust set of researchers that work on responsible AI”
Suha Can, at 14:41
As the team builds models, he says, it uses analytical linguists and machine learning models to check them. He names three things a model must not do.
“that they wouldn’t leak data, that they wouldn’t mislead our users by hallucinating”
Suha Can, at 15:03
The third is saying things that would be awful to say. A hallucination is an AI answer that presents false or misleading information as fact (Wikipedia).
“So typically, when I talk to other CISOs, they, of course, ask about compliance, but they’re also curious, so I end up explaining to them what we do with responsible AI.”
Suha Can, at 15:34
A few weeks after the episode aired, on September 4, 2024, Suha shared Grammarly’s white paper on how it approaches responsible AI, “The Responsible AI Advantage: Guidelines for Ethical Innovation.”
Suha Can on LinkedIn, September 4, 2024: Grammarly's white paper on how it builds responsible AI, with principles like transparency, fairness and user agency.
Suha Can · LinkedIn
Controls users can see
The third part of the trust story sits in the product itself.
“We also rely a lot on controls in the product.”
Suha Can, at 15:42
A company can tell people it cares about privacy and keeps a high security bar, he says, but it’s hard for them to look under the hood.
He gives two examples. Users can see the data Grammarly has stored against their user ID at any time. And they can decide whether to use a feature, with an explanation of the data behind it. He says Grammarly believes this is how it earns its users’ trust.
Grammarly published the same ideas in a December 2022 post: it doesn’t sell user data, it never runs in sensitive fields like password and credit card fields by default, and users can view all the data tied to their account (Grammarly).
The EU AI Act
Jon says he isn’t sure the security community knows what secure AI means yet, and thanks Suha for putting a team on the question. Sasha turns to regulation.
“Has EU AI regulations impacted the velocity or the roadmap of certain controls that are being implemented?”
Sasha Sinkevich, at 18:46
The Artificial Intelligence Act is the European Union’s regulation on AI. It entered into force on August 1, 2024, two weeks before the episode aired, and its provisions phase in over the following 6 to 36 months (Wikipedia). Suha welcomes it, and his reason is clarity.
“So I think typically, as a security professional, as a privacy professional, what I fear most is unknown and ambiguity, and there’s still a lot of that in the space.”
Suha Can, at 19:10
He says the act didn’t change how Grammarly handled EU users’ data, because Grammarly handled it the same way before.
“Some companies use data to improve their models, improve their products. Some companies hold on to data indefinitely. That’s not how Grammarly is operating.”
Suha Can, at 20:00
The General Data Protection Regulation already applied to those users. The EU adopted it on April 14, 2016, and it took effect on May 25, 2018 (Wikipedia).
Our internal compass
Sasha sums it up: strong foundations in data privacy and governance mean a new regulation doesn’t force sudden changes in how the product is built. Suha agrees. Whatever a regulation requires, he says, Grammarly’s internal compass is that its practices pass an internal test. The decodes split on the test’s name, so the transcript has its best guess in brackets. The principle under it is the customer first principle.
“Use customer data only to provide a service to the customer.”
Suha Can, at 21:35
Where does my data go?
Jon says he and Sasha have both heard the same request many times: AI services that run on-premises, or at least data ownership that stays on-premises while the services run in a managed cloud.
“Have you guys had a lot of those requests come in at Grammarly too? Or do you offer a single-tenant tier in your enterprise space?”
Jon McLachlan, at 22:10
Suha says the concern is real and bigger in the age of generative AI.
“When a company uses an AI tool internally, their data goes somewhere else.”
Suha Can, at 22:48
The concerns come down to two groups. The first is about storage and access.
“So in this space, it really boils down to concerns around retention. Do you store this data? How long do you store this data? Who can access this data?”
Suha Can, at 23:00
Regional data storage belongs to that group too. The second is about training.
“And the second piece is, do you use my data to train models? And if you do, is it to train models just for me, or is it to train models for all of your customers?”
Suha Can, at 23:15
A key the customer holds
Suha calls Grammarly’s answer a tiered model that keeps changing with customer feedback. Running everything themselves means the customer has to manage it, he says, and some customers don’t want to do that. They still want their data kept secure. The answer he describes is a key that’s controlled by the customer.
“So the customer hosts the key, gives us permission to the key so that we can use it to encrypt and decrypt data.”
Suha Can, at 24:12
“They can revoke our access to their data immediately by revoking our access to the key.”
Suha Can, at 24:25
They can also rotate their keys, he says, which gives them control over how their data is encrypted.
The general model is called bring your own encryption, or bring your own key. The customer manages its own keys, and if it deletes them, the encrypted data can’t be read again, a practice called crypto-shredding (Wikipedia). Grammarly announced Bring Your Own Key for enterprise customers on June 25, 2024, with session timeouts, custom roles and group-level security controls (Help Net Security).
Suha puts the customer-held key between two extremes: the cost of running a customer-guarded environment, and a setup where the customer has no visibility and no control. On training, he says Grammarly does not train on its enterprise customers’ data.
“On training, we absolutely do not train on the data of our enterprise customers, and it’s just not something we do.”
Suha Can, at 24:58
“I’ve seen application layer encryption, bring your own key style encryption, used in a past life to really open up the market for very similar types of products.”
Jon McLachlan, at 25:11
Three measures of success
Jon asks what success looks like for Grammarly’s security team, as far ahead as Suha wants to look. Suha starts with the work itself.
“And this space is full of hard and unsolved security problems.”
Suha Can, at 25:57
The first measure is to solve those problems and share how Grammarly brings safety into its models. If the team does that well, he says:
“We really made this technology safer for everybody.”
Suha Can, at 26:29
The second is to stay a user-centric and ethical AI company that keeps the user in the driver’s seat on what data to share.
The third is out of the team’s hands.
“So if you do those things well, and of course, the third is the market has to reward our approach, and it also has to be true for us to be successful.”
Suha Can, at 27:37
Hard, long days
Sasha asks for the best day at Grammarly so far. Suha can’t pick one, but he knows the kind.
“where the work is extremely hard, extremely long, and for a really good cause.”
Suha Can, at 28:34
“And doing a run like this, a really hard, long day with my team, colleagues, it’s just what gets me into my flow state.”
Suha Can, at 28:44
Jon likes a good flow state too, early in the morning or late at night, and sometimes both on the same day. Then he asks for the opposite: the single most challenging day of Suha’s career.
The blue team of the internet
Suha’s hardest days were at Microsoft, where he worked on security incident response.
“And also, you know, the adversaries that Microsoft deals with is that they are just truly outstanding, technically. It’s absolutely the best of the best.”
Suha Can, at 29:59
“And we kind of viewed ourselves there in the MSRC as kind of like the blue team of the internet.”
Suha Can, at 30:09
MSRC is the Microsoft Security Response Center. His LinkedIn describes the work: as part of the MSRC React team he found vulnerabilities in Microsoft products and did root cause analysis and variant investigation on reported ones, and later he led a product security engineering and research organization across the United States, Israel and China that responded to vulnerabilities and zero-day exploits (LinkedIn). A zero-day is a vulnerability unknown to anyone who could fix it (Wikipedia).
Some of those days, he says, aren’t wise to talk about in any context. Jon agrees: no violating NDAs.
The Shadow Brokers
One day he can talk about.
“It’s probably around 2017, there was this incident, the Shadow Brokers.”
Suha Can, at 30:30
“but some unknown parties came upon the internet with remarkably bad grammar, and they were teasing that they have a bunch of really, really good zero-days from the NSA that they will drop.”
Suha Can, at 30:37
The Shadow Brokers emerged in mid-2016 and first published on August 13, 2016. They leaked hacking tools, including zero-day exploits, from the Equation Group, widely suspected to be part of the NSA (Wikipedia). The teasing lasted months. Then the release came, and one exploit in it was EternalBlue.
“So it was real. They absolutely were authentic in their claims, the Shadow Brokers.”
Suha Can, at 31:06
EternalBlue was an NSA exploit built on a Windows zero-day that let attackers reach any number of computers on a network. The group released it publicly on April 14, 2017 (Wikipedia).
Patched before the drop
The hard part for Microsoft was the uncertainty. The team had a time limit and no list.
“But we didn’t know exactly which vulnerabilities these folks were going to release. There’s a lot of unknown and uncertainty there.”
Suha Can, at 31:46
The vulnerabilities in question were remote zero-days in Windows. Jon and Sasha point out how many people that touches.
“It’s not just individuals. Most of the world back then was running on Windows system. It still is.”
Sasha Sinkevich, at 32:07
So the team searched millions of lines of Windows code with little to go on, fixed as much as it could, and released the fixes. About a month later the release dropped.
“That day was a very stressful day for us.”
Suha Can, at 32:55
They reverse engineered the exploits and recognized each of them.
“Thankfully, we got it right. So we patched all of those vulnerabilities before the release happened. And obviously, the opposite would have been a disaster.”
Suha Can, at 33:11
Wikipedia gives the same gap. Microsoft had patched some of the Windows exploits in a security bulletin on March 14, 2017, a month before the April 14 leak (Wikipedia).
Then WannaCry
A couple of months later, WannaCry used one of those exploits, and Suha’s explanation is that not everyone patches. The WannaCry attack began on May 12, 2017. It spread through EternalBlue, mostly through organizations that hadn’t applied Microsoft’s patches or ran Windows versions past their end of life, and it hit more than 300,000 computers in 150 countries (Wikipedia). NotPetya used the same exploit on June 27, 2017 (Wikipedia).
Sasha sees a prioritization lesson in it.
“a lot of things are dependent on how well you prioritize this specific set of criticals and highs.”
Sasha Sinkevich, at 33:37
No researcher to call
Suha says that wasn’t the lesson of that day.
“This is not a prioritization exercise, because we definitely knew the importance of these vulnerabilities.”
Suha Can, at 34:14
Most vulnerability reports come from a researcher Microsoft can talk to. This time, he says, there was no researcher, and nobody was telling them anything. The Shadow Brokers and Microsoft had no channel between them.
The no-mystery principle
So what got the team through? Suha’s answer is a principle for working with that much ambiguity.
“the mindset has to be that there must be no mystery, no unanswered questions.”
Suha Can, at 34:50
The method is to argue against your own best guess until nothing is left to argue.
“You try to prove yourself wrong. You say, no, it can’t be it. What if it’s this?”
Suha Can, at 34:55
When every question has an answer, the answer that’s left is the one to trust.
“through answering all of the questions, only the truth remains”
Suha Can, at 35:15
He says he uses the principle with his teams. It means getting to the bottom of the problem, really understanding it, and trying to prove yourself wrong.
“And if you can’t, then I think you must be right.”
Suha Can, at 35:35
Sasha connects it to security and to life in general.
“In other words, we have to know what we’re trying to fix before we fix it.”
Sasha Sinkevich, at 35:56
The same idea shows up in what Suha said a year earlier about bug bounties. In a June 14, 2023 HackerOne write-up of his talk with HackerOne’s Alex Rice, he said, “Preemptive security is about working to disconfirm your beliefs.” He also said the main bug bounty metric he reports to the board is the number of unique researchers who reported at least one vulnerability in a quarter (HackerOne).
Believe in yourself more
Jon’s last question: would Suha meet his younger self, and what would he say? Suha says yes, just to freak the guy out. The advice is about confidence.
“I think it took me a very long time to get to a place where I feel I actually know what I’m doing.”
Suha Can, at 36:34
“So maybe what I will tell them, hey, believe in yourself more, take more risk.”
Suha Can, at 36:47
Jon puts it as stop having imposter syndrome, and Suha agrees.
Attackers and defenders with AI
Before Jon can close the episode, Sasha asks one more question: what Suha expects from product security and from AI in security over the next three to five years. Suha takes the AI half first.
“I think we will see a lot more misuse of advanced AI technology by adversaries, by malicious actors, to accelerate cybersecurity capabilities that they may have.”
Suha Can, at 37:38
He lists malware development, phishing and social engineering among the uses. On the defending side, he says AI is already showing up a little in security operations and incident response, and much more could happen there.
When a hallucination reaches the physical world
Then the risk side.
“we will probably run into cases where hallucinations actually have a real-world impact.”
Suha Can, at 38:22
“So something will happen in the physical world because of a hallucination”
Suha Can, at 38:27
He expects it in the news, and he expects it to push the industry toward resilience for AI and for critical infrastructure.
Misalignment
Further out, he names alignment.
“I do believe, with artificial general intelligence, the misalignment issues will come to bear, and we’ll see cases where AI will have its objectives not aligned with what the humans are trying to achieve.”
Suha Can, at 38:48
Alignment is the work of steering AI systems toward the goals people intend. A misaligned system pursues goals no one intended (Wikipedia).
Chief trust officer
The security half of his answer is about the job.
“I think the security role, and the role of the CISO in particular, is kind of becoming a lot more broad than it originally was.”
Suha Can, at 39:07
It started with corporate security and moved into product security, he says, and now companies want the CISO to lead the adoption of AI.
“A lot of companies are actually looking at the CISOs to lead the company in the adoption of AI and in the bringing of AI in the workplace.”
Suha Can, at 39:24
He believes the CISO is becoming a chief trust officer in every respect.
He likes the change. The role stays technical and strategic, he says, with a bigger seat at the table. Sasha says the job is about data: how you protect it, whether it goes into training, and who can reach it.
“Who has access to the data, whether it’s internal teams or it’s internal services?”
Sasha Sinkevich, at 40:10
“That was the first time I heard it posed as a chief trust officer. I like that a lot.”
Jon McLachlan, at 40:40
Suha says he knows a couple of people with the title already. Sasha closes by offering a follow-up episode about the future, and Suha says he likes talking about the future.
After the tape
Suha took the case on the road. Grammarly announced his session at the Gartner IT Symposium/Xpo on October 22, 2024, on a playbook for measuring the return on AI deployments (LinkedIn). In November 2024 he wrote in Grammarly’s “AI Responsibly” newsletter on how return on investment and responsible AI connect (LinkedIn).
Grammarly changed at the top. In December 2024 Grammarly agreed to buy Coda, and Coda’s CEO, Shishir Mehrotra, replaced Rahul Roy-Chowdhury as Grammarly’s CEO. In May 2025 it raised $1 billion in nondilutive funding from General Catalyst. In July 2025 it agreed to buy the email client Superhuman, and in October 2025 the parent company, Grammarly Inc., rebranded as Superhuman (Wikipedia).
Suha moved to DoorDash. He joined DoorDash as vice president and CISO in March 2025. His title now reads vice president and chief security officer, and he leads cybersecurity, governance, risk and compliance, and physical safety for the company and its subsidiaries (LinkedIn).
Suha Can on LinkedIn, March 2025: he joined DoorDash as VP and CISO.
Suha Can · March 2025
Grammarly has a new CISO. Grammarly’s trust center now names Giles Douglas as chief information security officer (Grammarly).
More from Grammarly on this show. Polina Morozov, then a security recruiter at Grammarly, was the guest on episode 46, two months before this one.
Did you catch it? A seven-question quiz
Seven questions from the episode. Each answer links the moment in the transcript.
Quiz · seven questions
Did you catch it?
Pick an answer to see whether you caught it, and where it is on the tape.
-
What did Suha find on the bottom floor of the coffee shop?
- One computer
- A shelf of Unix manuals
- A room full of hackers
a. "And at the bottom floor of the coffee shop, there was one computer" Transcript, 2:19.
-
What does Suha call his first job at Grammarly?
- Vice president of engineering
- The CISO job
- Chief privacy officer
b. "Job number one is the CISO job." Transcript, 4:41.
-
How does Suha describe scalable solutions?
- The whole job
- A waste of money for a startup
- Necessary but not sufficient
c. "and I truly believe that scalable solutions are necessary but not sufficient." Transcript, 7:16.
-
Which new certification was Grammarly pursuing when the episode was recorded?
- ISO 42001
- ISO 9001
- FedRAMP
a. "So in the compliance space, for example, we are really pursuing, I think we are one of the first to do that, the new ISO 42001" Transcript, 14:11.
-
What does Suha say Grammarly does not do with its enterprise customers' data?
- Encrypt it
- Train models on it
- Store it in the cloud
b. "On training, we absolutely do not train on the data of our enterprise customers, and it's just not something we do." Transcript, 24:58.
-
What did the Shadow Brokers say they would drop?
- Windows source code
- A list of stolen passwords
- Zero-days from the NSA
c. "they were teasing that they have a bunch of really, really good zero-days from the NSA that they will drop." Transcript, 30:43.
-
What does Suha's principle say there must be none of?
- Mystery
- Risk
- Meetings
a. "the mindset has to be that there must be no mystery, no unanswered questions." Transcript, 34:50.
Would an enterprise CISO trust your AI product?
Suha’s answers add up to a list of what an enterprise buyer checks before it lets an AI tool near its data. Eight questions on how many of them you can answer.
Scorecard · eight questions
Would an enterprise CISO trust your AI product?
Answer for the AI product you sell to businesses. If two answers fit, pick the lower one.
-
Can you tell a buyer, in writing, what customer data you store, for how long, and who can access it?
- No
- Partly
- Yes, on one page
"So in this space, it really boils down to concerns around retention. Do you store this data? How long do you store this data? Who can access this data?" See Where does my data go?
-
Do you train models on your enterprise customers' data?
- Yes, by default
- Only if they opt in
- No
"On training, we absolutely do not train on the data of our enterprise customers, and it's just not something we do." See A key the customer holds.
-
Can an enterprise customer hold its own encryption key, and revoke or rotate it?
- No
- On the roadmap
- Yes
"They can revoke our access to their data immediately by revoking our access to the key." See A key the customer holds.
-
Can users see the data you hold about them, and turn features on or off?
- Neither
- One of them
- Both
"We also rely a lot on controls in the product." See Controls users can see.
-
Does every model go through a check for data leaks, hallucination and harmful output before it reaches users?
- No
- Some models
- Every model
"that they wouldn't leak data, that they wouldn't mislead our users by hallucinating" See A dedicated responsible AI team.
-
Do you hold SOC 2 Type 2 or ISO 27001, with a plan for ISO 42001?
- Neither
- One, no plan for 42001
- Yes, and a plan for 42001
"We are also quite aggressive in pursuing new certifications, especially in the realm of AI and privacy, because especially with AI, there is a lot of nervousness in the industry." See One of the ways we communicate.
-
Does your security team build controls with the engineers, or advise and walk away?
- Advise only
- Some of each
- Build with them
"So my teams don't operate in a way where we advise teams what to do and then we move away." See I didn't know security was this.
-
Could you explain your data practices to a regulator in one sentence?
- No
- Roughly
- Yes
"Use customer data only to provide a service to the customer." See Our internal compass.
A buyer will ask about retention, access and training before anything else. Write a one-page answer, and decide where you stand on training before they ask.
The basics are in place. Next, give enterprise customers a key they hold, and put a check on every model before it reaches users.
You can answer what Suha says every enterprise asks. Keep the answers current as your models change.
How many machines are still unpatched on exploit day?
Microsoft patched before the Shadow Brokers release, and WannaCry still spread a month after it, through machines that hadn’t been patched. This calculator shows how many of yours would still be open when an exploit goes public, and when a worm arrives.
Calculator
How many machines are still unpatched on exploit day?
Four numbers: your machines, the share of unpatched machines you patch each week, and the days from the patch to a public exploit and to a worm.
With 10,000 machines and 30% of the unpatched ones patched each week, 2,061 are still open when the exploit goes public after 31 days, and 495, or 5% of them, when the worm arrives after 59 days.
The 31 and 59 days are the gaps from Microsoft's March 14, 2017 security bulletin to the Shadow Brokers release on April 14 and to WannaCry on May 12, 2017. The 10,000 machines and the 30% weekly rate are examples we picked. Put in your own numbers.
From a coffee shop in the 1990s to DoorDash in 2025
Everything on this page sits on one timeline. Filter it by thread.
Timeline · the 1990s to 2025
From a coffee shop in the 1990s to DoorDash in 2025
Filter by thread. Suha's dates are from the tape and his LinkedIn, and the rest are from the public record.
- The coffee shop
One computer, the IRC, and a Unix security group (on this show).
- IBM
Enterprise IT security intern, to October 2003 (LinkedIn).
- AirTies Wireless Networks
Research and development engineer on embedded Linux routers, owning the firewall and the product's security, to 2007 (LinkedIn).
- Microsoft Security Response Center
Security engineer on the MSRC React team, to January 2013 (LinkedIn, on this show).
- Grammarly
First release, from founders Alex Shevchenko, Max Lytvyn and Dmytro Lider (Wikipedia, on this show).
- Amazon Web Services
Security engineer and penetration tester, to August 2013 (LinkedIn).
- Back to the MSRC
Principal security engineering manager, leading product security engineering and research, to December 2018 (LinkedIn).
- GDPR adopted
The EU's data protection regulation, in effect from May 25, 2018 (Wikipedia, on this show).
- The Shadow Brokers
The group's first publication (Wikipedia, on this show).
- Microsoft's patches
A security bulletin patches some of the Windows exploits, a month before the leak (Wikipedia, on this show).
- EternalBlue goes public
The Shadow Brokers release it with other NSA tools (Wikipedia, on this show).
- WannaCry
The ransomware worm spreads through EternalBlue to more than 300,000 computers in 150 countries (Wikipedia, on this show).
- NotPetya
The same exploit helps carry out the attack (Wikipedia).
- A browser extension bug
Tavis Ormandy of Google finds that Grammarly's extension exposed authentication tokens to websites, and Grammarly ships a fix within hours (Wikipedia).
- A bug bounty
Grammarly launches its program on HackerOne, with a $100,000 reward for the first hacker to reach a specific document on its server (Wikipedia, on this show).
- Amazon Payments Security
Suha founds the team and grows it from 1 to more than 100 engineers, then leads security engineering for Amazon Payments from March 2021 (LinkedIn, on this show).
- Grammarly
CISO and vice president of engineering, to February 2025 (LinkedIn, on this show).
- A user-first post
Grammarly says more than 50,000 teams use Grammarly Business and that it doesn't sell user data (Grammarly, on this show).
- Generative AI
Grammarly launches a product built on large language models (Wikipedia).
- Preemptive security
HackerOne writes up Suha's talk on bug bounties and board metrics (HackerOne, on this show).
- Episode 46
Polina Morozov, then a security recruiter at Grammarly, on the show (episode 46).
- Bring Your Own Key
Announced for enterprise customers, with session timeouts and custom roles (Help Net Security, on this show).
- The EU AI Act
Enters into force, with provisions phasing in over 6 to 36 months (Wikipedia, on this show).
- Episode 52 airs
This conversation (Podbean).
- The responsible AI white paper
Suha shares Grammarly's guidelines for responsible AI (LinkedIn, on this show).
- Gartner IT Symposium/Xpo
Grammarly announces Suha's session on measuring the return on AI deployments (LinkedIn, on this show).
- Coda and a new CEO
Grammarly agrees to buy Coda, and Shishir Mehrotra becomes CEO (Wikipedia, on this show).
- DoorDash
Suha joins as vice president and CISO (LinkedIn, on this show).
- ISO/IEC 42001
Grammarly announces its certification (Business Wire, on this show).
- Superhuman
Grammarly agrees to buy the email client (Wikipedia, on this show).
- A new name
The parent company, Grammarly Inc., rebrands as Superhuman (Wikipedia, on this show).
What to do this week
Five things, in the order of the episode.
- Write down where customer data goes. Put on one page what you store, for how long, who can access it, where it lives, and whether it trains any model. Suha says every enterprise asks those questions. Take the scorecard with your product lead.
- Decide who holds the key. If buyers ask for on-premises, a key the customer holds may be the middle ground Suha describes. Our Product Security team can help you design it.
- Check every model before it ships. Test for data leaks, hallucination and output you would never want to ship, and write down who signs off. ISO 42001 turns that habit into a management system an auditor can certify.
- Time your patch race. Run the calculator with your own fleet and patch rate, then measure how long a critical patch takes to reach most of your machines.
- Get a second pair of eyes. A vCISO can own the plan, and you can book 15 minutes with us.