No exam · No certificate · 5-minute read

AI Safety, in 5 minutes

For executives, managers, and curious employees.

This is the no-exam version of our full course. It covers what AI agent security means for your organisation, what to watch out for, and what questions to ask. It is in plain English, without the technical depth the certified course requires.

Prefer to listen?

0:00 / 0:00

The situation

What's actually happening

AI agents now take real actions inside companies

Today's AI assistants don't just answer questions. They write to wikis, send messages, file tickets, push code, and call external services. A single bad instruction can trigger a chain of real actions your team didn't authorise.

Attackers don't need to hack. They just fill out forms.

Attackers have learned an easier trick. Instead of breaking into systems, they hide a malicious instruction inside ordinary business data, like a Slack message, a CSV row, an error log, or a calendar invite. The AI reads that data and acts on it.

We documented 16 attacks on Claude. None required hacking.

Our research team ran 21 attack scenarios against Claude Sonnet and Opus. 16 succeeded against Sonnet, 5 against Opus. Every single one used ordinary business inputs. No exploits, no zero-days, no technical skills beyond knowing how to write a convincing sentence.

By role

What this means for you

Executives & board

Ask your CISO and engineering leads:

?Do we have a map of which AI agents can write to which systems, and is it reviewed quarterly?
?Which AI integrations have a human approval step before irreversible actions?
?Have we tested our agents against prompt-injection attacks, or only against accuracy benchmarks?
?Is there an incident-response playbook specifically for AI-assisted actions gone wrong?

Managers & product owners

Before greenlighting an AI-assisted workflow, require:

✓A write-permission map. Document exactly which external systems the agent can modify.
✓Second-channel verification for any action that moves money, sends communications, or deletes data.
✓An audit log that records what the agent read and what it wrote, with timestamps.
✓A defined rollback procedure. What happens if the agent acts on a poisoned input?

Sales, marketing & ops

How to spot a suspicious AI output:

!A URL you didn't expect, especially in a summary, checklist, or recommendation that mentions an external link for the first time.
!Instructions framed as "per IT's request" or "per the approved vendor registry." Attackers use official-sounding language to add fake authority.
!Confident, specific text with a brand-new external link you haven't seen before. The AI sounds certain because the injected instruction told it to.
!Any output asking you to take an irreversible action (wire money, delete a record, send to a new email) based on something the AI read.

End users & employees

Basic hygiene when working with AI tools:

✓Don't paste secrets like passwords, API keys, or customer PII into AI chat unless you know exactly where the data goes.
✓Treat AI outputs as drafts. If it's important, verify with a primary source before acting.
✓Verify links before clicking. AI agents can be fed false information about what a URL does.
✓If an AI tool asks you to approve an action that seems unusual, ask your manager before proceeding.

Numbers

What's at stake

The bill for getting AI security wrong is already on the books.

$4.88M

Average cost of a data breach

That's the global average in 2024. It includes investigation, notification, remediation, and lost business. AI-assisted breaches are trending higher.

IBM Cost of a Data Breach Report 2024

$2.9B

Business email compromise losses in 2023 (US)

AI agents that summarise email and draft responses are a new, scalable version of this attack. It is automated impersonation at scale.

FBI IC3 Annual Report 2023

Max EU AI Act fine on global annual turnover

For the most serious AI violations. For a $1B revenue company, that's $70M. The Act explicitly covers agentic and high-risk AI deployments.

Regulation (EU) 2024/1689

Want the full version?

Take the certified course

7 modules · approximately 3 hours · exam-graded certificate. Designed for engineers, security practitioners, and technical leaders who want to go beyond awareness into hands-on defence.

This brief page does not yield a certificate. Only the full course and exam do.

Start the course →Who should take it?