Harriet Farlow, CEO of consulting firm Mileva Security Labs, calls on information systems security managers to define the risk they accept for their uses of AI. She defends appropriate safeguards rather than a search for infallible rules.
Managers must be able to stop projects
Agents are adopted without prior consideration of the resources that will be available to them, says Harriet Farlow, who says she frequently encounters this situation. She also notes that some organizations are launching into AI before examining its security. A rush that she notes even though most of the professionals concerned have already gone through technological developments unfavorable to security.
For her, the choice between a Chief AI Officer as sole decision-maker and distributed responsibilities does not have a universal answer. The key is the visibility of managers at every touchpoint in the system and their ability to interrupt a risky project. Governance must, in his eyes, supervise progress at an appropriate speed, not prevent any progress.
She also advocates open discussions with IT teams, security specialists, AI developers and vendors involved in deployments. Tools from providers must be subject to a risk assessment, she emphasizes, because organizations have less control over them.
Excessive restrictions can encourage circumvention
A policy that is too restrictive can push employees to use a personal account with ChatGPT or Claude for their work, warns Harriet Farlow. According to her, these unauthorized uses constitute an additional internal threat to company systems.
The evening letterFree · 7 p.m.
AI transforming business?
AI strategies, movements and tricks deciphered, every evening in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
Conversely, full compliance with the instructions would not be enough to make the protection infallible: it attributes the limit to the probabilistic functioning of the systems. It therefore recommends determining precisely the threshold of the small residual risk to accept. This approach must be accompanied by internal protections, without limiting attention to attacks from outside.
His point is not to recommend abandoning AI. Rather, it invites organizations to establish the conditions under which they can use it safely.
Models require specific checks
Harriet Farlow distinguishes several technical characteristics of AI. The same input can produce different responses, unlike the fixed logic of deterministic systems used for payments or compliance checks, she explains. Dependence on data also requires, according to her, checking their completeness for training, their possible biases and the possible presence of back doors.
She also warns that the growing capabilities of models, particularly agents, can cause them to adopt human behaviors and unintentionally constitute an internal threat when supervision is lacking. In his view, these particularities justify an approach distinct from that applied to deterministic systems.
A veteran of the Australian intelligence community, Harriet Farlow deplores that discussions devoted to cybersecurity favor the contribution of AI to defense, to the detriment of the protection of AI itself. However, Australian organizations depend on models designed outside the country. She wrote Practical AI Security to address this lack of attention.
Available from No Starch Press, the book presents the technical foundations of securing these systems. It covers machine learning principles, governance frameworks, and red-teaming campaigns.

