SECURITYWEEK’S Kevin Townsend argues that while current AI cannot be trusted, its transformative potential means guardrails and secure practices are essential if we are to rely on it in the future. He highlights core problems: AI’s dependence on probability rather than truth, hallucinations, bias, and a tendency toward sycophancy, all of which can mislead users and misinform decisions.
The piece notes that a 2023 paper by Ilia Shumailov, published in Nature in 2024, introduced the idea of model collapse, describing how models trained on data generated by earlier generations can degrade over time. It also points to practical risks, such as the example of an acting director of CISA uploading sensitive data to a public chatGPT, which underscores the need for pre- and post-release guardrails and safer data handling.
As DeepKeep and other AI-defence firms push “brain rewiring” and external guardrails, Townsend concludes that we cannot trust current AI, but we cannot afford not to use it, with corporate users bearing responsibility for securing implementations. The article was published on 9 April 2026.