The CISO and AI: risk is no longer measured asset by asset

For years, the chief information security officer (CISO) has fought to sit on the management committee and talk about business, risk and resilience. They have succeeded. And right now they face the most uncomfortable exam: artificial intelligence is not just a new technology that has to be inventoried, it is a factor that alters the rules by which risk has been calculated up to now. The reflection is signed by Arancha Jiménez, vice president of the Cybersecurity Services and Products Line at Atos in Spain and Portugal, in an article published in Computerworld.

The central idea deserves pause. The problem is not that an attacker uses autonomous agents, but that AI reduces the limitations that have historically held the adversary back. Two hypotheses on which many risk models have been built fall away: that the attacker's effort is limited and that they will only exploit what yields a benefit that clearly outweighs the cost. With AI, the time and resources required are compressed to speeds that current defenses find hard to match.

From the asset list to the relationship map

The article gives an example that is very recognizable to any IT team: a medium vulnerability in an exposed service, a role with excessive privileges and a misconfigured segmentation rule. Seen separately, the three findings seem manageable. Connected, they can form a viable path to critical assets. That is where the shift in approach lies: the severity of a vulnerability must be complemented by its real exposure, the context, the effective possibility of exploitation and its impact within an attack chain.

The CISO who continues to address risks from an individual perspective will be defending a list of assets, while the adversary will operate on a map of relationships.

This is not a new idea, the author acknowledges. What until recently was complex was putting it into practice almost immediately. And this is where AI plays in your favor: agents can correlate findings from multiple tools, analyze them across thousands of assets, identify invisible relationships, prioritize systemic risks and reduce remediation, containment and response times. AI versus AI, in short.

What it means for an SME

Translated to a small or medium-sized company, the message is not "hire an AI-powered SOC." It is more down-to-earth: stop looking at findings in silos. Many SMEs live with an incomplete inventory, accounts with privileges inherited from an emergency and firewall rules that no one has reviewed since implementation. Each piece, in isolation, alarms no one. Together, they draw a path.

The author herself points out two conditions without which this model does not work. The first is data quality: reliable inventories, well-identified dependencies and criticalities and a real understanding of exposure. Without that base, AI will amplify errors. The second is the governance and security of the AI systems themselves: a poorly governed defensive agent, without least privilege, monitoring, traceability and explicit limits on autonomy, becomes a new risk vector.

Practical recommendations for the IT team

1. Organize the inventory before automating anything. If you do not know what you have, what depends on what and what is critical, any layer of AI on top will only accelerate wrong conclusions.
2. Review accumulated privileges. Excessive roles are the usual fuel for attack chains. A quarterly review costs little and avoids scares.
3. Look at segmentation with an attacker's eyes. A misconfigured rule is not a minor finding: it is a bridge between two zones that should have been separated.
4. Prioritize by path, not only by severity. Ask what combination of weaknesses would allow reaching sensitive data, even if each piece has an average score.
5. Govern any agent you deploy. Least privilege, traceability and limits on autonomy from day one. It is the same discipline you would apply to any service account.

This approach connects with what we have already seen on the blog: when a single poorly protected credential opens the door to dozens of environments, as in the case of ConnectWise and the 173 Microsoft 365 tenants, the risk is not in the credential, it is in the relationships it enables. Something similar happens with the arrival of offensive AI tools to more teams: Anthropic opening its cyber AI to verified teams expands the available arsenal, also for those looking for shortcuts.

The article's conclusion is a useful warning: the biggest risks are rarely hidden in the obvious, but in what we have not yet learned to observe. For an SME, starting to observe the connections —identities, configurations, dependencies— is probably the security investment with the best return before adding more tools.

Source: Computerworld. Analysis and adaptation: ForgeNEX.