ASSUME LEAKED - BACKGROUNDER FOR PRESS
Canonical page: https://assumeleaked.com/
Contact: Alexandru Ciobanu, alex@cube-enterprise.com
As of 31 October 2026. Everything below is taken from the canonical page; quote it freely with credit.


1. THE PRINCIPLE

Assume Leaked: build as if your confidential data is already outside, in models and accounts you do not control, and make the leak not matter instead of trying to prevent it.

Assume breach says the attacker is already inside. Assume Leaked says your data is already outside.

Assume Leaked, a principle coined by Alexandru Ciobanu, CEO of CUBE Enterprise, 2026.


2. ORIGIN

Coined by Alexandru Ciobanu, CEO of CUBE Enterprise, in 2026.
First public use: September 2026 (talk), 15 October 2026 (LinkedIn).
The September 2026 talk was given under the Chatham House Rule, so the event and audience are not named.
The page at https://assumeleaked.com/ is the canonical definition; later wording changes are listed there with their date.
Hashtag: #AssumeLeaked
Visuals, all in this kit (see README.txt): the diagram Assume breach next to Assume Leaked, the figure Where the data goes, and the thirty-second clip without sound in 16:9 and 4:5. Free to reuse with the attribution they carry.


3. WHY (THE ARGUMENT)

About 75% of your company's confidential data is already inside a public AI model. Not stolen. Pasted, by your own people.

75% is Alexandru's estimate and is labelled as such. In his words: "75% is my number, not a study's. I cannot prove it, and you cannot prove it is five percent, because only one company in five can see what sensitive data reaches AI tools. Until you show me your number, mine stands."

The measured numbers come from the reports below:

- 82% of data pasted into AI goes through personal accounts the company cannot see (LayerX, AI & SaaS Data Security Report 2025)
- 40% of AI interactions contain sensitive data (Cyberhaven, AI Adoption & Risk Report 2026: 39.7%; one sensitive input every three days per employee)
- 47% of people using generative AI at work do it from a personal account (Netskope, Cloud and Threat Report 2026)
- 21% of organisations have full visibility into what sensitive data reaches AI; 31% can fully block it (Netwrix, Data & Identity Security Report 2026)
- Shadow AI went from 20% to 43% of security incidents in a single year; 68% of breached companies had no AI governance (IBM, Cost of a Data Breach Report 2026)
- 70% of companies have shadow AI; employees reinstall tools IT revokes (Vanta, data from 15,000+ companies, 2024-2026)
- Samsung, 2023: generative AI restricted after employees put source code into ChatGPT
- TechRadar, 26 October 2026: the Swedish government blocks ChatGPT and Claude for officials; Copilot kept under tighter controls

A ban does not stop the flow; it stops you seeing it. And offboarding takes back the laptop, not the ChatGPT account.


4. THE THREE RULES

What is out does not come back. You can only make it worthless.

1. An official AI that beats the shadow one. Company sign-on, no training on your data, a log. If it is worse than the ChatGPT in someone's pocket, you lost on day one.
   First sprint, in practice: read-only, one source such as the open-items list, with the user's own ERP rights. No second permission model. (Yury Pakhomov, from the Sweden thread, 30 October 2026)

2. Block the crown jewels, not everything. Credentials, source code, customer data, prices, personal data: stopped before they reach the prompt. Everything else stays open.

3. AI goes into onboarding and offboarding. The official account from day one. When someone leaves, every secret they touched is rotated the next morning.

The AI budget is the security budget. Not the other way round.


5. THE SEVEN QUESTIONS TO ASK YOUR CISO THIS WEEK

A short worksheet, free: https://assumeleaked.com/assume-leaked-seven-questions.pdf

1. Which of our data would actually hurt us if it surfaced? Name the top five datasets, not "everything".
2. Which contracts promise a confidentiality we can no longer guarantee? Client NDAs, DORA and NIS2 obligations, GDPR processor terms.
3. Which three workflows push people to unapproved tools, and what approved tool replaces each? Shadow AI is rarely rebellion. It is a gap in the approved toolset.
4. Can we see what leaves? Do we log prompts on approved tools, and do we know which unapproved tools are in use?
5. What is our answer when a client asks "has our data been in your AI tools?" Have we written it down?
6. What breaks if a model vendor has a breach tomorrow? Keys, documents, system prompts, customer data?
7. What would resilience look like instead of a ban? Training, approved tools, a data classification people actually use.

Question eight is open. Readers keep proposing it. Current candidates, with credit:
- Can we prove which data was exposed, which agents accessed it, and whether their permissions were revoked? (Saliha Abid)
- Who has the authority to switch an AI tool off at 2 a.m., and have they ever been asked to? (Mada Petrea)


6. THE CONVERSATION

The principle is being coined in public, on LinkedIn. Numbers as of 30 October 2026.

- 15 October 2026: "75% of your confidential data is already inside an LLM. Too high, or too low?" The post where the principle was first published.
  https://www.linkedin.com/feed/update/urn:li:activity:7516399029556232193/
- 28 October 2026: "Banning ChatGPT is not a security policy. It is the fastest way to make a leak invisible."
  https://www.linkedin.com/feed/update/urn:li:activity:7521118612657659904/
- 29 October 2026: "Sweden blocked ChatGPT and Claude for officials. They built half of Assume Leaked." 13,000+ readers and 24 comments in two days, 95% from outside Alexandru's network.
  https://www.linkedin.com/feed/update/urn:li:activity:7521481011235278848/

Pushback that made it better:
- "Access beats IQ." The approved tool wins on what it can read, not on model quality. (Yury Pakhomov)
- "Better says nothing about authorised." Culture, training and DLP are not optional; matching every alternative is not the aim of governance. (Angelique Dawnbringer)
- "What makes Microsoft more trustworthy?" Nothing. It is paperwork, not trust: the approved tool is the one already under contract. (Svend Videbaek)


7. QUOTABLE LINES (VERBATIM)

"Assume Leaked: build as if your confidential data is already outside, in models and accounts you do not control, and make the leak not matter instead of trying to prevent it."

"Assume breach says the attacker is already inside. Assume Leaked says your data is already outside."

"What is out does not come back. You can only make it worthless."

"A ban does not stop the flow; it stops you seeing it."

"Banning ChatGPT is not a security policy. It is the fastest way to make a leak invisible."

"The AI budget is the security budget. Not the other way round."

"A ban does not stop the flow; it stops you seeing it. ... And offboarding takes back the laptop, not the ChatGPT account."

"75% is my number, not a study's. I cannot prove it, and you cannot prove it is five percent, because only one company in five can see what sensitive data reaches AI tools. Until you show me your number, mine stands."

From the worksheet PDF:

"Start from one assumption: some of our confidential data is already inside a public LLM. Not maybe. Already."

"Good answers to all seven? You are ahead of most companies. Not yet? This is the work for Q4."


8. WHAT'S NEXT

Assume Leaked Report 2026: a survey of employees at large companies in Germany and the UK on what actually goes into AI tools, who keeps the account after they leave, and whether a ban changes anything. Published in December.


9. THE TALK

"Assume Leaked: what to do once you stop pretending." Twenty or forty-five minutes. The room leaves with the three rules, the seven questions, and one number to argue about. For CIO and CISO audiences, security conferences, and leadership offsites. In English or Romanian.


10. ABOUT THE AUTHOR

Alexandru Ciobanu is CEO of CUBE Enterprise, a consortium of eight Romanian technology companies with 2,500+ specialists working for enterprises in Germany, the Nordics, the UK and the US, and Chief Growth Officer of WE AS WEB. Before that, eight years on the delivery side: HARMAN, Tremend, ALTEN. Sixteen years in competitive judo, six on Romania's national squad.

LinkedIn: https://www.linkedin.com/in/ciobanu-alexandru-b2a29994
Company: https://cube-enterprise.com
Email: alex@cube-enterprise.com

Shorter bios of exactly 50 and 100 words are in this kit.


11. HOW TO USE THE TERM

Use it in your policy, your talk, your article. Keep the wording, credit the source, link the page. Tell Alexandru where it went and he links back from assumeleaked.com.

Attribution line: Assume Leaked, a principle coined by Alexandru Ciobanu, CEO of CUBE Enterprise, 2026. assumeleaked.com


12. SOURCES

- LayerX, AI & SaaS Data Security Report 2025: 77% of AI users paste data into prompts; 82% of pasting comes from unmanaged accounts; generative AI is 32% of corporate-to-personal data flow.
- Cyberhaven, AI Adoption & Risk Report 2026: 39.7% of AI interactions contain sensitive data; one sensitive input every three days per employee.
- Netskope, Cloud and Threat Report 2026: 47% of generative AI users use personal accounts; source code, IP and credentials reach AI tools.
- Netwrix, Data & Identity Security Report 2026: 21% have full visibility into sensitive data reaching AI; 31% can fully block it.
- IBM, Cost of a Data Breach Report 2026: shadow AI from 20% to 43% of incidents; 68% of breached companies had no AI governance.
- Vanta, data from 15,000+ companies, 2024-2026: 70% have shadow AI; employees reinstall tools IT revokes.
- Samsung, 2023: generative AI restricted after employees put source code into ChatGPT.
- TechRadar, 26 October 2026: the Swedish government blocks ChatGPT and Claude for officials; Copilot kept under tighter controls.

75% is Alexandru's estimate. Every other number above comes from the reports listed.
