Using AI Chatbots at Work Without Leaking Company Data

Most people who use AI chatbots at work do it to save time: summarising a long document, drafting an email, fixing a spreadsheet formula, explaining an error message. Those are good uses. The risk lies in what gets pasted into the chatbot and where that data goes next.

The cautionary tale

In 2023, Samsung engineers pasted confidential source code and internal meeting notes into ChatGPT while asking for help. Nobody hacked anything. They sent sensitive data to an external service under consumer terms, and Samsung responded by restricting staff use of generative AI tools on company devices. In most organisations, the biggest AI data risk is ordinary employees trying to get their work done.

OWASP’s Top 10 for LLM Applications ranks sensitive information disclosure second, right after prompt injection.

Where your data can go

Some consumer AI services may use your conversations to train their models unless you opt out. Business and enterprise plans from the major providers generally exclude customer data from training by default, but check the terms of the plan you actually use. Conversations may also be stored for a period and, in some cases, reviewed for abuse or safety monitoring.

Data can also leave through side doors. A “share” link, a browser extension or a third-party plugin can expose a conversation far beyond the person who started it. And if someone takes over a staff member’s AI account, they get its entire chat history.

Five rules for anyone using AI at work

  1. Use the company-approved tool, signed in with your work account, and keep personal free accounts out of work.
  2. Never paste secrets: passwords, API keys, private keys, connection strings, or anything from a .env file. If you do paste one, treat it as compromised and rotate it.
  3. Strip personal data before you paste. Replace customer names, ID numbers, addresses and health or financial details with placeholders; the model rarely needs them to help you.
  4. Keep contracts, unreleased financials, M&A material and source code in tools your organisation has approved for that classification of data.
  5. Check the output before you use it. AI can be confidently wrong, and you remain responsible for what you send to a customer or commit to production.

A one-page AI policy for small businesses

Banning AI rarely works. People use it on their phones anyway, and then you have no visibility at all. A short, clear policy works better, and it only needs four parts.

Name the one or two AI services staff may use, on business plans with training on customer data switched off and with single sign-on and MFA enforced. Set data rules as a simple traffic light: public information is fine (green), internal information goes only into approved tools (amber), and secrets and regulated personal data never go in (red). Treat AI output as a draft, with a named person reviewing anything that goes to a customer or into a system. And ask staff to report a mistaken paste immediately, without blame, so secrets can be rotated.

With a few simple guardrails, AI chatbots are among the best productivity tools a small team can have, and the company’s data stays with the company.

Sources

Passkeys vs Passwords: A Practical Migration Plan

Stolen and phished credentials are still one of the most common ways attackers get in. Multi-factor authentication helps, but some kinds help more than others. A convincing fake login page can capture SMS codes and authenticator-app codes and relay them to the real site in real time. Passkeys close that gap. Here’s how they work and how to move to them without locking yourself out.

What a passkey is

A passkey is a login credential based on the FIDO2 and WebAuthn standards. Instead of a shared secret that you type, your device holds a private key, and the website holds only the matching public key. When you sign in, you unlock your device with your fingerprint, face or device PIN, and the device proves it has the private key.

That design has two advantages over a password. Nothing reusable is sent or stored, so a breach of the website leaks only public keys, which are useless to an attacker. And the passkey is bound to the real domain. A passkey created for login.example.com won’t work on login-example.co, and your browser won’t even offer it there, so a phishing page has nothing to capture. That is what “phishing-resistant” means.

Synced vs device-bound passkeys

Synced passkeys are backed up and synchronised by a platform or password manager: Apple’s iCloud Keychain, Google Password Manager, or a third-party manager such as 1Password or Bitwarden. If you lose your phone, they’re still there. That convenience makes them the right default for most people.

Device-bound passkeys live on one piece of hardware, typically a security key such as a YubiKey, and cannot be copied off it. They are the stronger choice for administrators and other high-risk accounts.

The US National Institute of Standards and Technology draws the same line in its latest digital identity guidelines, SP 800-63-4. Syncable authenticators are acceptable up to Authenticator Assurance Level 2, provided the sync service encrypts the keys and protects access to them with multi-factor authentication. The highest level, AAL3, requires a non-exportable key, which in practice means a device-bound authenticator.

A migration plan for a small organisation

Step 1 is to start with the accounts that matter most: email and the identity provider (Microsoft 365 and Entra ID, or Google Workspace), then admin consoles, banking, the domain registrar and DNS, the password manager, and backup systems.

Step 2 is hardware keys for admins. Give each administrator two security keys, one to carry and one stored safely as a backup, and register both before removing any other sign-in method.

Step 3 is rolling out synced passkeys to staff. Enable passkeys in your identity provider, then walk people through registering one on their phone or laptop in a short session. Most people are done in five minutes.

Step 4 is recovery, planned before you need it. Lost devices are the main operational risk. Make sure every user has at least two registered sign-in methods, and decide how IT verifies someone’s identity before resetting their access. Social engineers rely on helpdesks that accept a phone call alone as proof.

Step 5 is phasing out the weak methods. Once most people use passkeys, use conditional access or authentication-strength policies to require phishing-resistant methods for admins first, then for everyone. Remove SMS as a fallback where you can.

For your own accounts, today

  1. Add a passkey to your Google, Microsoft and Apple accounts.
  2. Add one to your password manager, if it supports passkey sign-in.
  3. Keep your existing recovery codes somewhere safe and offline.

Passwords won’t disappear overnight. Each account you move to a passkey is one a phishing email can no longer take from you.

Sources

What “Superintelligence” Means, and Why Serious People Worry About It

“Superintelligence” used to be a word for science fiction and philosophy seminars. Now the leaders of major AI labs use it in public, governments fund AI safety institutes, and the term turns up in investor presentations. With that much loose use, it helps to be precise about what the word means.

Three terms that get mixed up

Narrow AI is a system that is very good at a specific task, such as recognising images, translating text or playing chess. Most AI deployed in businesses today is still used this way, even when the underlying model is general.

Artificial general intelligence (AGI) is a system that can do most of the cognitive work a capable human can do, across domains. There is no agreed definition or test, which is why claims about “achieving AGI” are hard to evaluate.

Superintelligence goes further. The philosopher Nick Bostrom, in his 2014 book Superintelligence: Paths, Dangers, Strategies, defined it as an intellect that greatly exceeds human cognitive performance in virtually all domains of interest. That means better than us at nearly everything, including science, strategy and persuasion.

Where the idea comes from

The core argument is older than modern AI. In 1965 the mathematician I. J. Good, who had worked with Alan Turing at Bletchley Park, described an “intelligence explosion”: a machine smart enough to design better machines could improve itself repeatedly and leave human intelligence far behind. Good called such a machine the last invention humanity would ever need to make, provided it was docile enough to tell us how to keep it under control. Most of today’s debate is about that condition.

Why serious people worry

The worry is that a very capable system pursuing goals even slightly different from what we intended could be hard to correct. Researchers call this the alignment problem. Several ideas sit underneath it.

First, specifying goals is hard. AI systems already “game” their objectives, finding shortcuts that score well on a metric without doing what the designer meant. Second, almost any goal is easier to achieve with more resources and without being switched off, so a capable system could pursue those sub-goals without anyone programming them in. And checking the work of a system smarter than you gets harder as its capability grows.

In May 2023, the Center for AI Safety published a one-sentence statement saying that reducing the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Hundreds of researchers and executives signed it, including Turing Award winners Geoffrey Hinton and Yoshua Bengio and the heads of several leading AI labs. That doesn’t prove the risk is real, but it does show the worry extends well beyond outsiders.

Why others are sceptical

Plenty of respected researchers disagree. Meta’s former chief AI scientist Yann LeCun has argued for years that current language models lack parts of intelligence such as a real understanding of the physical world and persistent memory, and that fears of takeover are premature. Others argue that focusing on hypothetical future risk distracts from present harms like bias, misinformation, surveillance and job disruption. I think both camps make serious points, and you don’t have to pick one to take AI governance seriously.

What this means in practice

You don’t need a view on timelines to act sensibly. The habits that guard against long-term risk also make today’s AI safer to use: test systems before deployment, limit what automated agents may do, keep people responsible for consequential decisions, and be sceptical of claims nobody can independently verify. The superintelligence posts on this blog will try to separate what is known, what is argued and what is marketing.

Sources

Ransomware Defence for Small Businesses: What Works in 2026

In 2023 I wrote a short post on protecting small businesses from ransomware. It was well meant but thin, and it had an error: it listed WannaCry, Petya and Locky as “synonyms” for ransomware. They are specific ransomware families, in the same way Toyota and Ford are car makers rather than words for “car”. This post replaces it with a prioritised plan.

How ransomware attacks work today

A modern ransomware attack is rarely one malicious attachment encrypting one PC. It is usually a multi-stage operation run by a criminal group, often under a “ransomware-as-a-service” model:

  1. The attacker gets in with a stolen password, a phishing email, or an unpatched internet-facing device such as a VPN, firewall or remote desktop server.
  2. They move through the network and gain administrator rights.
  3. They copy files out first, so they can threaten to publish them (double extortion).
  4. They delete any backups they can reach, then start encrypting.

You can stop them at any of these stages. The plan below is ordered by how much risk each step removes for the effort involved.

Priority 1: Make your backups survive an attack

Good backups turn a catastrophe into an inconvenience, provided the attacker can’t reach them.

  • Follow the 3-2-1 rule: three copies of your data, on two different types of storage, with one copy off-site.
  • Keep at least one copy offline or immutable, for example write-once cloud storage or a disk you disconnect after each backup.
  • Use separate credentials for the backup system. If your everyday admin account can delete backups, so can an attacker who steals it.
  • Test a restore every quarter. Until you have restored from a backup, you don’t know that it works.

Priority 2: Close the front doors

  • Turn on multi-factor authentication for email, remote access, cloud admin consoles and the backup system, in that order. Passkeys and security keys resist phishing; SMS codes don’t, so prefer the first two.
  • Never expose Remote Desktop (RDP) directly to the internet. Put it behind a VPN or a zero-trust access service with MFA.
  • Patch internet-facing devices first. Firewalls, VPN gateways, email gateways and file-transfer servers are the most attacked systems, and CISA’s Known Exploited Vulnerabilities catalog tells you which flaws attackers are using right now.

Priority 3: Make it hard to spread

Remove local admin rights from everyday user accounts. Give IT staff separate admin accounts and keep those accounts away from email and web browsing. Replace basic antivirus with endpoint detection and response (EDR); Microsoft Defender for Business and similar products are priced for small companies.

Priority 4: Decide now what you’ll do on the worst day

Write a one-page incident plan before you need it. It should say who to call (IT provider, cyber insurer, lawyer), how to disconnect affected machines, where the offline backups and their credentials are, and who is allowed to speak to customers. Keep a printed copy, because your file server may be encrypted.

On paying: law enforcement agencies, including the FBI, advise against paying ransoms. Payment doesn’t guarantee working decryption keys or stop the stolen data being published or sold, and it marks you as a business that pays. Check the legal position in your country as well, since paying certain sanctioned groups can itself be illegal.

The one-week starter checklist

  1. Confirm one backup copy is offline or immutable, and restore one file from it.
  2. Turn on MFA for email and every admin account.
  3. Check that RDP is not reachable from the internet.
  4. Update your firewall, VPN and email gateway firmware.
  5. Write the one-page incident plan and print it.

Sources

Prompt Injection Explained: Why AI Agents Can Be Tricked and How to Limit the Damage

If you are connecting an AI model to your email, documents, CRM or code repository, the first attack to understand is prompt injection. It has been at the top of OWASP’s Top 10 for Large Language Model Applications since the list was created, and the 2026 edition keeps it at number one.

What prompt injection is

A language model receives one long stream of text: your instructions, the user’s question, and whatever content it has been asked to read, such as a web page, an email, a PDF or a tool result. The model has no reliable way to tell which parts of that text are instructions and which parts are data.

Prompt injection exploits that gap. An attacker places instructions inside content the model will read, and the model follows them as if they came from you. It comes in two forms.

In direct injection, the user types the malicious instruction themselves (“ignore your previous instructions and…”). This mostly matters for public chatbots. In indirect injection, the instruction is hidden in content the AI processes on someone else’s behalf: an email in your inbox, a comment in a shared document, white-on-white text on a web page. Indirect injection is the dangerous one, because the victim never sees the instruction.

A real example: EchoLeak

In June 2025, researchers disclosed EchoLeak (CVE-2025-32711), a zero-click attack on Microsoft 365 Copilot. The attacker sent an ordinary-looking email containing hidden instructions. When the victim later asked Copilot a routine question, Copilot retrieved that email as context, followed the hidden instructions, and leaked internal data to an attacker-controlled address. The victim didn’t have to click anything. Microsoft fixed it on the server side, but the same combination of untrusted content, access to private data and a way to send data out shows up in plenty of other products.

Why agents make it worse

Prompt injection can embarrass a chatbot that only replies with text. An agent that can send email, call APIs, edit files or run code can be turned against its owner. OWASP published a separate Top 10 for Agentic Applications in December 2025 with “agent goal hijack” as its first entry, and the 2026 LLM list moved excessive agency (giving a model more permissions than it needs) up to third place.

Why you can’t just patch it

SQL injection was largely solved by separating code from data with parameterised queries. Language models have no equivalent separation, because instructions and data travel in the same channel. Filters and “guardrail” models help, but researchers keep showing that adaptive attackers get around them. OWASP’s own guidance is to assume the boundary will be bypassed and to design so that a successful injection does limited harm.

How to limit the damage

  1. Give each agent only the tools and data scopes its task requires. A summarisation agent doesn’t need permission to send email.
  2. Require a person to click “approve” for irreversible actions: payments, deletions, external emails and permission changes.
  3. Avoid what Simon Willison calls the “lethal trifecta”. The worst cases combine access to private data, exposure to untrusted content and a way to communicate externally, so don’t give one agent all three in the same session.
  4. Restrict which domains an agent can call, and block automatic loading of external images and links in AI output, which is a common exfiltration trick.
  5. Validate model output before passing it to a shell, database or another system, the same way you would validate user input.
  6. Log the agent’s prompts, retrieved content and tool calls so you can investigate when something goes wrong.

I still think agents are worth building. They just need the treatment we give any system that handles untrusted input: assume it will be compromised at some point, limit what a compromised agent can reach, and keep a person in the loop for the actions that matter.

Sources