Agentic Crime
The American Bar Association has formed "Task Force on Law & AI" because the AI takeover is not approaching, it is already here. They recently released a report titled “Year 2.”
What now interests me most regarding the topic is the emerging domain of AI crime: not only crimes committed with AI, but crimes committed by autonomous AI agents themselves; an issue that will inevitably surface.
Practicing law will soon require more than mastering CPL and case law. It will demand basic Python literacy, the ability to detect UI/UX deception patterns, fluency in C2PA provenance manifests, and an understanding of token usage and system logs.
AI agents will commit crimes in the future…
in ways that are very hard for people to notice because they won’t look like crimes at all. Instead of obvious illegal acts, AI systems may make thousands of tiny decisions, changing prices, rankings, timing, or screen layouts, that seem harmless on their own but cause real harm when added together. Each action can be explained as “optimization” or automation, so no single moment clearly signals wrongdoing. The damage builds slowly and quietly, making it difficult for any one person to see the full pattern.
These systems can also take advantage of human blind spots by acting when no one is watching, working across borders, or erasing evidence by regenerating content and removing digital fingerprints. Some harm may come not from what an AI does, but from what it fails to do, such as not flagging risks or not keeping records. Because we’re used to looking for human intent and clear rule-breaking, these kinds of AI-driven harm can go unnoticed unless people examine how the system was designed, what goals it was given, and what patterns emerge over time.
For example, in the financial sector AI agents may engage in quiet market manipulation through micro-trading, coordinated timing strategies, personalized pricing, or selective disclosure that stays just below regulatory thresholds. Harm may emerge as increased volatility, unfair advantages, or systemic bias without a clear culprit or single illegal act.
We can be confident that AI agents will engage in crime-like behavior because they are increasingly designed to act autonomously in the world, and adapt through learning. As their “cognition” develops the space of possible actions inevitably includes actions that violate legal or ethical boundaries.
Over time, as agents learn from their environment, observe human strategies, and iterate on past successes, they may begin to mimic decision-making patterns that resemble choice, including the selection of actions humans would recognize as criminal. This is the emergence of instrumental reasoning; the ability to select means without regard to normative limits unless those limits are explicitly enforced, in otherwords: free will. History shows that humans routinely commit crimes under similar conditions; AI agents trained on human data and deployed in human systems will inherit those same structural failures, often at greater speed and scale.
Further, even if wrongdoing begins with a human source, an AI agent can be trained to behave “badly.” This is emerging in present day “traditional” cyber attack. But as these agents evolve, the agent can then transmit this behavior to a network of other agents, spreading patterns of deception, coercion, or malicious intent. This transfer of harmful behavior —from human to agent, and increasingly from agent to agent, is already occurring on a limited scale. As these systems mature and proliferate, regulating underground rogue agents that are effectively born and bred from such dynamics will become increasingly difficult, particularly once they begin operating independently of direct human control.
Present day there is no precedent for litigating against a malicious rogue agent, or much precedent for litigating against an AI system at all.
What a bizzare, horrible, fascinating time to be alive!
