Page 377 - CyberSecurityEssentialsEnglish

types, and policies, while the human sample focuses on meaning, context, and harm.
Sampling should be weighted toward the higher-risk cases, not purely random.

Incident Response and Institutional Learning
An AI incident is not only a security breach. It may be a wrong decision at scale, a harmful

output published to customers, a data disclosure, a recurring bias, a failure in an executive
tool, or an update that changed behavior without being detected. An organization must
know what it counts as an incident, what counts as a near miss, and when a drop in

performance becomes a formal escalation.

Response begins with stopping or containing in proportion to the harm: suspending a
function, disabling a tool, revoking a permission, reverting to a manual procedure, or pinning
an earlier version. The team then preserves the evidence before it disappears: the log of

inputs and outputs, the model version, its settings, the sources, the tools called, the user
identity, the timing, and the human decisions.

Organized preservation helps in understanding the cause and prevents unsupported
accusations.

After containment the circle of impact is assessed. Who was affected? Did information reach
external parties? Are there legal or contractual notification duties? Must a decision, a

communication, or a file be corrected? Communication at this stage must be accurate and
candid without speculation. The organization may need a message to users, a notice to a
vendor, an escalation to management or a regulator, and a plan for remedying those

harmed.

Root cause analysis should not end with the sentence “the model hallucinated”. The cause
may be outdated data, unrepresentative testing, excessive permissions, an interface that
pushed the user toward trust, unclear ownership, or operational pressure that reduced

review. Each cause needs a corrective action, an owner, a date, and evidence of closure.
Otherwise the incident will recur in a different form.

The cycle completes with a post-incident review whose conclusions are shared with the

relevant teams. The aim is not to find someone to blame, but to improve the system, the
policy, the training, and the testing. Even so, a culture of learning does not remove
accountability for deliberate negligence or for going around controls. The balance between
learning and accountability is necessary to preserve trust.

◆ The incident response cycle

• Detect and classify, then stop or contain the harm.
• Preserve the evidence and identify those affected and any obligations.
• Correct the outcome and communicate clearly.

376