Solving for a new exposure from AI in one area — say, bias — could generate exposure in a new one like cybersecurity. Hugh Mulligan of consultancy S-RM explores the interconnected nature of AI governance and risk.
Every American company is feeling the pressure to incorporate AI into nearly every function or risk losing its competitive edge. In many respects, that pressure is justified. In terms of speed and cost, for many companies, AI tools are no longer just hype.
But the governance conversation around AI adoption has a structural problem: It treats bias, privacy and cybersecurity as separate risks to be managed in sequence. In practice, though, they interact, and the mitigation an organization adopts for one routinely changes its exposure to the others, often without anyone noticing. Organizations that miss this end up exposed across multiple risks at once, sometimes precisely because of the steps they took to address just one.
The fix becomes the vulnerability
Bias offers the clearest illustration of this dynamic. Large language models don’t reason the way humans do. In their current state, LLMs convert language into numerical representations, learn statistical patterns from enormous volumes of text and generate responses by predicting what sequence of words is most likely to come next. This is closer in principle to predictive text than to a human mind with judgment or intent. As a result, a model’s “understanding” is only as good as the data it absorbed. If that data contains skewed viewpoints, historical prejudice or uneven coverage of certain groups, the model internalizes those patterns and reproduces them at scale, often with a tone of confidence that makes bias harder to detect.
Suppose an organization detects model drift, outputs that have fallen out of sync with the real world. The standard response is to broaden the training data, pulling in additional geographies, demographics and edge cases and to set up continuous data refresh. That reduces bias, but it quietly enlarges the attack surface: each new pipeline, vendor, API and connector creates additional places where sensitive data is stored or transmitted and additional systems and people granted access to it.
In effect, the bias fix has created a cybersecurity problem.
The same dynamic plays out between privacy and security. To manage privacy risk responsibly, organizations add controls: logging, retention schedules, consent management, data subject access request workflows, audit trails, provenance tracking. These are good practices, but each one creates new targets: log databases that may contain personally identifiable information, credentials or prompts; admin consoles and review queues that need to be secured; third-party relationships with ID verification vendors, ticketing systems and data discovery platforms, each of which is its own potential entry point.
The organization built a privacy compliance program and inadvertently handed attackers a new map of its infrastructure.
Cybercriminals, meanwhile, are not standing still. Generative AI is being used to scale capabilities that once required highly resourced and coordinated groups, automating reconnaissance, producing convincing social-engineering lures and generating new malware variants faster than defenders can triage.
How powerful that scaling has become is visible most clearly in defender-side experiments. Anthropic’s Mythos model, first released only to a small circle of partners under Project Glasswing, has reportedly surfaced thousands of previously unknown software vulnerabilities in weeks of testing. Anthropic gated that capability precisely because of what it could do in the wrong hands. On the criminal side, the same kind of capability has no equivalent gate.
That compounds an asymmetry that’s always been central to cyber defense: an attacker needs one path to succeed, while a defender has to close all of them. AI accelerates both, but it does more for the side that only needs to win once.
A patchwork without a compass
If entanglement of risks is the first problem, the US regulatory environment is the second, and it compounds the first. Many American organizations assume that operating outside the stricter EU framework means they have more room to maneuver. What they actually have is more regimes to answer to, with no central one to satisfy.
The EU AI Act establishes a unified, tiered framework that classifies AI systems by risk level and assigns obligations accordingly, prioritizing legal certainty and rights protection at the cost of some flexibility. The UK has taken a more “pro-innovation” path, distributing AI oversight across existing regulators rather than creating a new statutory regime, on the view that adaptability serves both growth and adoption better than a single rulebook. Neither approach is perfect, but both give organizations a coherent framework to work within.
Canaries in the Coal Mine: Law’s Crashout Over AI Is Coming for Everyone
Shadow AI & lack of proper guardrails are problematic combination
Read moreDetailsThe US offers neither. American companies operate across three overlapping and often conflicting layers: a federal framework that is less a single law than a patchwork of agency guidance, executive actions and existing statutes stretched to cover AI; state-level rules that vary widely and sometimes contradict each other; and industry-specific regulations layered over both. The federal layer is also actively unstable: The current administration’s deregulatory posture has pulled hard against the more prescriptive framework of its predecessor, leaving organizations to track federal expectations against a moving target while state regimes continue to harden. The fragmentation is itself a governance risk because there is no single standard of care to build toward, no unified test of compliance to pass.
Consider a national retailer rolling out an AI hiring tool that screens resumes and analyzes video interviews. Federally, existing anti-discrimination law still applies; an algorithmic decision doesn’t provide immunity. In New York City, using an automated employment decision tool triggers requirements for bias audits and candidate notices. In Illinois, AI-reviewed video interviews require specific notice and consent practices, a 30-day deletion obligation if an applicant requests it and strict written policies around any biometric data collected. In Colorado, where a new AI law took effect in February, that same hiring tool is classified as a high-risk system and the deployer is required to use reasonable care to avoid algorithmic discrimination, conduct annual impact assessments and notify candidates that AI was used. These requirements don’t point in the same direction. An organization that builds its compliance posture around one of them is probably exposed under another.
American organizations shouldn’t assume their AI systems will navigate this landscape on their own, and they shouldn’t assume that informal monitoring — employees periodically consulting an AI chatbot about regulatory questions — constitutes a compliance program. It doesn’t.
What structured governance requires
The natural institutional response to a problem with several faces is to divide it: bias to data science, privacy to legal, security to IT. That structure feels logical, and in many organizations, it maps onto existing functions. But it produces predictable failures. Each team manages its own exposure without visibility into how its mitigations affect the others, and the resulting compliance record — when a regulator, insurer or board asks how a specific deployment was assessed — is scattered across functions, inconsistent in depth and impossible to reconstruct chronologically.
The organizations managing AI risk effectively do something different. They treat each AI deployment as the unit of risk, sitting at the intersection of bias, privacy and security simultaneously and assess it accordingly. This isn’t a wholly new idea; regulated financial institutions have governed model risk this way for years, and GDPR’s data protection impact assessments work on similar logic. What’s new is applying that discipline across the broader risk surface — bias, security and compliance alongside privacy — that AI deployments now create.
That starts with a feature-level inventory: what does this system touch, what data does it ingest, what decisions does it influence and where does it connect to other systems and third parties? The inventory has to extend beyond the deployment itself. Ten low-risk deployments sharing a common pipeline can produce aggregate exposure that no single deployment carries on its own, which means risk has to be tracked at the platform level as well. And because most enterprise AI sits on top of third-party foundation models, vector stores and pipeline tooling, governance also has to reach into the supply chain — what the vendor tested for, what the deployer tested for, and where the seam between them is.
From that inventory, deployments are tiered by exposure. A customer-facing AI that ingests personal data and influences access decisions is a different risk profile from an internal tool that summarizes meeting notes, and resources should be concentrated where the exposure is highest. Most deployments will sit in lower tiers needing only light-touch assessment, freeing capacity for the high-tier minority that warrants rigorous testing, frequent review, integration scrutiny and clear escalation. Across all tiers, the process should generate documentation as it goes — contemporaneous records of what was assessed, what was found, what was decided and why — rather than retrospective compliance paperwork.
Human oversight in this model is not a hand on every output. At scale, per-decision review is infeasible and would defeat much of the point of adopting AI in the first place. The right level is the system: defining the boundaries within which a deployment is allowed to operate, monitoring whether outputs stay inside them, escalating exceptions, owning incident response and periodically revalidating that the system is still doing what it was deployed to do. Organizations that build that into the operational layer — with defined responsibilities, documented decisions and regular stress-testing of the program against scenarios it would actually have to handle (a model behavior shift, a vendor compromise, a regulator’s question, a coordinated attack on the integration) — will have something to show for it. Those that appoint a nominal AI lead and run an annual review will not.
The case for acting now
The point of detailing these risks isn’t to encourage hysteria or skepticism around AI. AI, on balance, is a net positive, and organizations that treat adoption as inherently dangerous will find themselves at a real competitive disadvantage. But adoption without an expert, systematic approach to governance is itself a risk, one that becomes harder to correct the longer it goes unaddressed.
The entanglement of bias, privacy and regulatory risk means that organizations can’t patch their way out of exposure incrementally. They need a view of the whole system. That requires mapping how AI deployments interact across risk domains, building tiered governance frameworks and producing the kind of defensible documentation that holds up under scrutiny.
Hugh Mulligan is an associate director in cyber risk and governance at S-RM, a global intelligence and cybersecurity consultancy. He previously served in a variety of roles at PwC UK. 
