Free learning.
Three short primers in plain English. Free to read and share.
Prompt injection, explained.
An AI model reads its instructions and the content it is given as one stream of text, and it cannot reliably tell the two apart. Prompt injection is text that the model treats as an instruction when it should have been treated as data.
- Direct: someone types instructions that try to override the rules, such as "ignore your previous instructions".
- Indirect: the instructions are hidden in something the AI reads, such as a web page, a document or an email. They need not be visible to a person.
It matters most when the AI can act: send email, read files, call tools. Injected text can then make it act for the attacker.
What helps
- Least privilege: give the AI access to only what the task needs, so a manipulated assistant can do little harm.
- Human in the loop: a person approves high-risk actions, such as sending money or deleting data.
- Treat outside content as untrusted: keep it separate from instructions and label where it came from.
- Test it: try the attacks before an attacker does. OWASP notes there may be no fool-proof prevention, so layered controls matter.
Sources: OWASP LLM01:2025 Prompt Injection (opens in a new tab); NCSC, Prompt injection is not SQL injection (opens in a new tab); MITRE ATLAS (opens in a new tab).
Shadow AI: five questions to ask.
Shadow AI is AI used for work outside an organization's approved systems and processes. It is a form of shadow IT.
- Which AI tools are people using for work, approved or not, including browser extensions and AI features inside other apps?
- What data has gone into them? Client, customer, personal or confidential data is the first concern.
- Work account or personal account? Consumer services may store, retain or use what is entered to improve the service unless specific privacy controls are in place.
- Is anything connected? An AI tool linked to email, files or other systems can reach whatever that link allows.
- Is there an approved alternative that meets people's needs? Shadow AI is unlikely to disappear completely; the aim is to reduce the risk, with clear rules and approved tools that meet people's needs.
Source: NCSC, The hidden risks of shadow AI (opens in a new tab).
Deepfake payment fraud: a verification checklist.
The FBI warns that criminals use AI-generated voices and video to impersonate people, including executives in live video calls, to get money moved.
- Never approve a payment or a change of bank details on the strength of one call, email or video meeting.
- Call back on a number you already hold, not one given in the request.
- Agree a verification phrase, or require a second approver, for large or unusual payments. The FBI suggests a secret word or phrase; the same idea works inside an organization.
- Treat urgency and secrecy as warning signs, not reasons to skip a check.
- If it happens, report it quickly. In the United States, report to the FBI's Internet Crime Complaint Center.
Source: FBI IC3 public service announcement on generative AI and financial fraud (opens in a new tab).