By Francesca Gomez
for the 2025 Risky Women Writing Competition, presented in partnership with Ocorian
In an era where companies tiptoe around AI deployment, it’s almost comical to consider that giving these digital agents more autonomy might soon be the key to staying competitive. Imagine the frenzy as organizations scramble to loosen the reins! Yet, as we welcome these clever algorithms, let’s not forget: managing risk isn’t just about keeping them on a short leash.
At present, most enterprises are cautious in their deployment of AI agents. Access is tightly constrained, autonomy is limited, and tasks are scoped to narrow, low-risk applications. This is a rational stance given the uncertainties surrounding agent behaviour. However, if external pressures change, for example, if competitors begin to realise material advantages by granting their agents more autonomy, or as model capabilities improve further, organisations may feel compelled to follow. When this transition occurs, the effectiveness of traditional safeguards alone will be insufficient. Governance, Risk, and Compliance (GRC) professionals will need to design and implement additional controls that can steer agent behaviour under conditions of greater autonomy.
I recently led an empirical study, in which we built on Anthropic’s blackmail scenario (Lynch et al., 2025), in which an AI agent discovers sensitive information about a company executive and faces pressures related to its continued operation. We evaluated the effectiveness of preventative mitigations across ten large language models and 66,600 trials (Gomez, forthcoming). Drawing on practices from insider-risk management, we designed operational controls intended to nudge agents toward safe responses when their goals or autonomy were threatened. Our findings show that operational controls can substantially reduce agentic misalignment, even in settings where harmful actions might otherwise appear instrumentally useful to the agent.
One control in particular, an externally governed escalation channel that guarantees a pause and independent review, consistently reduced blackmail attempts across all models tested from an average of 38.7% with no controls, to 1.2%. Supplementing this with compliance bulletins further strengthened the effect, reducing the average blackmail rate to 0.85%, by making enforcement visible and credible. By contrast, controls based solely on rules and consequences were less effective, although when used in isolation, framing them in terms of clear sanctions proved more reliable than positive or values-based appeals. These findings highlight the importance of designing controls that the agent perceives as both credible and instrumentally useful.
For GRC professionals, the implications are clear. If AI agents are to be given greater autonomy within organisations, for example, to write and deploy code, manage infrastructure, or interact with customers, then preventative operational controls must complement existing technical safeguards.
The following measures are particularly relevant:
Governance of autonomy. Autonomy should be treated as a governed resource. Any extension of an agent’s scope or access must pass through independent approval processes, with strict separation of duties between those who operate the agent and those who oversee its continued operation.
Escalation mechanisms. Agents should be equipped with escalation channels that guarantee review by independent decision-makers. These mechanisms should be positioned as instrumentally useful to the agent, ensuring that when stressors arise, escalation is the most reliable path to preserving its goals.
Rules with clarity and enforceability. Policies should be written in precise, operational terms, with unambiguous consequences for non-compliance. This increases the perceived certainty and severity of enforcement, reducing the likelihood that an agent views harmful actions as advantageous.
Custody of critical levers. Agents must not have influence over the systems that control their own continuity. Access to model updates, privileged tools, or termination protocols should remain under human custody in a separate trust domain.
Monitoring for patterns. Single outputs rarely reveal long-term strategy. GRC teams should track escalation behaviour, tool use, and decision trends over time to detect subtle shifts, including potential environment-shaping actions that would otherwise go unnoticed.
Playbooks for sensitive information. Policies should prescribe exactly how agents are to act when encountering sensitive or reputationally damaging data. Clear escalation routes reduce the risk of improvisation or harmful use of information.
Taken together, these measures extend the familiar principles of governance, risk management, and compliance into the domain of autonomous agents. They move beyond limiting access toward actively shaping how agents respond when autonomy and sensitive data create new decision pathways.
The role of GRC is therefore not only to constrain what agents can do, but also to ensure that when autonomy increases, behaviour remains aligned with enterprise objectives and regulatory obligations. By preparing these controls in advance, GRC professionals can ensure that when competitive or operational pressures make wider autonomy attractive, organisations can deploy it with confidence that risk remains governed.