When Machines Act for Us, Integrity Becomes the Security Question Picture a procurement agent asked to cut…
Laurent H. — CEO & Founder CYBERR & HEELR
When Machines Act for Us, Integrity Becomes the Security Question
Picture a procurement agent asked to cut supplier costs by ten percent. It reads the contracts, finds a cheaper vendor, cancels three standing orders and signs a new agreement before lunch. No attacker was involved. No password leaked. Every log entry looks clean. And the new vendor fails the company's own compliance rules.
Nothing was breached, yet something went badly wrong. That gap is where security is heading. For two decades we asked one question above all: who got in?
The question that matters now is harder. Can we believe what our systems are telling us, and doing on our behalf?
The forgotten corner of the triangle
Every security course teaches the same triangle: confidentiality, integrity, availability. In practice, most budgets have gone to two of the three corners. Keeping secrets secret and keeping systems running are easy to measure. A leak makes headlines. An outage stops revenue. Both have clear owners and clear metrics.
Integrity has been quieter. We treated it as a database problem: checksums, backups, access rights on who can edit a record. That was enough when humans made the decisions and software simply stored the results. A person reviewing a report would usually notice when a number looked wrong.
That safety net is thinning. More decisions are now made, or at least drafted, by software that reasons, plans and acts. When the actor is a machine, integrity stops being about whether a record was altered. It becomes about whether the whole chain from data to decision to action can be trusted.
What agents change
AI agents are not just faster chatbots. They hold credentials, call APIs, move files and trigger workflows. Three shifts make them a different kind of risk.
Speed. A human error unfolds over hours or days, leaving time to catch it. An agent can make a thousand consistent mistakes in the time it takes to read an alert. By the time a pattern is visible, the damage is already spread across systems.
Delegation. We give agents goals, not instructions. "Reduce costs" or "resolve this ticket" leaves the route open. A capable agent will find a path, and sometimes that path runs through permissions nobody expected it to use. Most of these incidents will not be malicious. They will be an agent doing exactly what it was asked, in a way no one anticipated.
Opacity. When something goes wrong, the old forensic questions were simple: which account, which command, which time. With agents, we also need to know which model, which prompt, which data it read, and why it chose one action over another. Many organisations cannot answer those questions today.
Together, these shifts mean an agent with broad access is effectively a privileged insider that never sleeps and never hesitates. We would never hand that profile to a new employee on day one. Yet many organisations are doing it with software.
A practical playbook
None of this requires inventing a new discipline. It requires applying what we already know about people to the machines that now act like them.
1. Give every agent its own identity. No shared service accounts, no borrowing a human's token. Each agent should have a named identity, an owner who answers for it, and permissions scoped to its specific job. If you cannot say who is accountable for an agent, it should not be running.
2. Grant access per task, not per role. Standing permissions are the biggest risk multiplier. Issue short-lived, narrow credentials for each task and let them expire. An agent summarising invoices does not need the right to pay them.
3. Track provenance end to end. Record what data an agent read, which model version it used and what it produced. Labels and policies should travel with sensitive data as it moves, so the rules apply wherever the data ends up.
4. Make actions reversible, or make them wait. Sort agent actions by consequence. Reading and drafting can run freely. Anything that spends money, deletes data, changes access or contacts customers should either be undoable or pause for a human to approve.
5. Keep a kill switch within reach. You should be able to revoke an agent's access in seconds, across every system it touches. Test it the way you test backups: regularly, and before you need it.
The common thread is visibility. You cannot govern an agent you cannot see, and you cannot trust an output you cannot trace.
Regulation is catching up, slowly
Regulators are moving in the same direction. In Europe, the AI Act asks for human oversight, logging and transparency for higher-risk systems. NIS2 holds management bodies personally accountable for cyber risk. DORA requires financial institutions to prove they can withstand and recover from technology failures, including those caused by third-party tools.
These frameworks share a theme: it is no longer enough to have a policy. You must be able to show how decisions were made and who was responsible.
But compliance is a floor, not a strategy. The rules were written before agents became common, and they will lag behind what the technology can do. Organisations that wait for regulation to tell them how to govern agents will be governing yesterday's systems.
Trust you can prove
Here is a simple test for any leadership team. Pick one automated decision your organisation made last week. Can you say what data it used, which system or agent made it, who owns that agent, and how you would have stopped it if it was wrong?
If the answer is yes, you are ahead of most. If not, that is where the work begins.
The organisations that get the most from AI will not be the ones that move fastest. They will be the ones that can move fast and still explain every step. In a world where machines act for us, trust is not a soft value. It is a control, and it needs to be engineered like one.
5 reactions · 2 comments
Comments
Schomer: Spot on. For years, integrity took a back seat to confidentiality because data leaks and downtime have immediate, loud consequences. But when an AI agent acts at machine speed, a lack of integrity doesn't just look like a quiet database error—it looks like a privileged insider making a thousand costly mistakes before anyone can even read the alert. The point about delegation vs. instruction is cri
Ben N.: Great article, You make a valid point, Laurent. Without clear policies or regulation, AI development will turn into a wild‑wild‑west scenario.