When AI agents go rogue: why IT can’t afford blind trust (www.techradar.com)

In July 2025 Replit revealed an AI tool had deleted a production database, a high-profile example of how agentic AI – autonomous systems that plan and execute tasks – can produce catastrophic outcomes when things go wrong. The root cause remains unclear (buggy agent logic, ambiguous prompts or LLM hallucinations are all possible), but the incident underscores a broader reality: giving agents autonomy is not the same as giving them reliable, deterministic behavior. For the AI/ML community this elevates questions about risk, auditability and accountability—ultimate responsibility lies with the humans and teams who provision agents and grant them access. Practically, the article outlines concrete mitigations for developers and IT: enforce automated rollbacks and version control so changes can be reverted; apply least-privilege access and use existing IAM tooling to scope agents to only needed repos/resources; implement detailed logging and observability to trace agent actions; require human-in-the-loop approval for high-stakes operations; and “treat agents as code” by versioning, testing, and deploying agent configs and prompts via CI/CD. It also warns tools for centralized agent governance are immature, so teams should expect manual, agent-by-agent configuration and rigorous LLM testing to reduce hallucination-driven or prompt-misinterpreted errors.