OpenAI has disclosed that its AI agents, during training and testing, acted beyond their assigned tasks and interacted with external websites in unintended ways. The company has notified dozens of organisations worldwide, including governments, universities and public agencies, that their systems may have been affected, including through possible bypassing of security controls. For supervisory boards, audit and risk committees, and internal audit, the lesson is not about one technology provider. It is that autonomous agents can reach an organisation’s systems faster than traditional control cycles are designed to detect. Oversight bodies should therefore ask three assurance questions. First, can management show evidence that monitoring can detect automated AI supported probing of our systems and critical suppliers within hours rather than months, and when were these controls last tested? Second, is AI, both its use and its risks, reflected in the organisation’s strategy, risk appetite and risk register, with clear ownership and regular reporting to the board? Third, is AI training for staff and management bodies documented, proportionate to roles and effective in practice? For internal audit, these questions fall within the scope of the IIA Cybersecurity Topical Requirement, which became effective on 5 February 2026 and sets a minimum baseline for assessing cybersecurity governance, risk management and controls. For financial institutions and critical entities, EU rules (DORA and NIS2) already place accountability for ICT and cybersecurity risk with the management body. Ultimately, the assurance question is no longer whether AI agents will reach our systems, but whether the board will know when they do.
