RESPONSIBLE AI
Managing frontier risks and agentic systems As AI systems grow more sophisticated, the nature of risk becomes more complex. Google’ s report places significant emphasis on“ frontier risks,” including cyber threats, harmful manipulation and the challenges posed by autonomous agents. These are problems that arise from increased capability and autonomy, meaning they are qualitatively different to existing issues.
One of the report’ s most notable contributions is the introduction of“ Critical Capability Levels” – thresholds at which a model’ s abilities may pose severe risks if left unchecked. By defining these levels, Google attempts to translate abstract concerns into actionable criteria. This approach aligns with risk management practices in other high-stakes domains, though its success will depend on how transparently these thresholds are defined and applied.
Elsewhere, discussion around harmful manipulation is particularly significant. The report acknowledges that advanced AI systems may be capable of influencing users in subtle yet powerful ways, raising concerns about autonomy and consent. While mitigation strategies are outlined, including alignment techniques and monitoring systems, the broader societal implications of such capabilities remain an open question.
Agentic AI represents another major focus. As systems begin to act on behalf of users, performing tasks such as browsing, planning and decision-making,
WITH AI AGENTS THE AIM IS TO SHORTEN DISCOVERY CYCLES, REDUCE COSTS AND OPEN UP NEW AREAS OF INVESTIGATION THAT WERE PREVIOUSLY TOO COMPLEX TO TACKLE AT SCALE
118 May 2026