Who answers when an agent acts?

Banks triage the bots

Pindrop identified 727,000 AI agents or bots in a two-month period, 94% acting legitimately and 6% malicious, and says agents are entering banking systems faster than banks can respond. Banks need new policies to distinguish legitimate customer-authorized bots from malicious actors. The company recommends a three-tiered routing strategy, green, gray and red, to manage risk based on the agent's identity and intent. m:…0044

Banks are conducting day-after scenario planning to secure accounts against AI agent errors, and Pindrop notes that checking a balance and changing a payout address carry different risk levels, so routing must weigh the agent's intent and action, not just its identity. The underlying problem is new: a caller that sounds human, acts on a customer's behalf, and may or may not be authorized, and the bank has to decide in real time which one it is. m:…0044

Disputes go to an AI arbitrator

General Legal has updated its website Terms of Service to include a mutual arbitration clause using the AI-powered service DecisionLayer. The company says the change follows a demand letter from a cookie troll, and DecisionLayer claims to resolve disputes 30 times cheaper and 10 times faster than standard processes. m:…0265

The provision is mutual, allowing either side to elect it, and DecisionLayer requires a human arbitrator to conduct a final review before an award issues. It applies only to website terms and does not affect existing client engagement letters. It is a narrow change, but it is a precedent worth watching: a legal services company routing disputes to an AI-powered process, and writing that choice into the terms users accept. m:…0265

From paper policy to enforced policy

ModelOp has announced Full-Stack AI Policy Enforcement integrated with NVIDIA OpenShell to manage AI policy across process, execution, operating system and network levels. The company says the platform addresses runtime policy violations by agents that bypass traditional lifecycle governance, and that integrating with OpenShell provides kernel-level controls over files, processes and network connections, with NVIDIA OpenShell running agents in a sandbox that denies access by default under kernel-level controls. m:…2216

The thread connecting banks, terms of service and kernel controls is accountability. Agents act, and when they act wrongly someone has to answer: the bank that accepted the call, the company whose terms routed the dispute, or the enterprise whose enforcement failed to stop the action. The organizations moving first are not waiting for a settled legal standard; they are building the routing, the terms and the enforcement now, and the standard will likely be written from what they build. m:…0044 m:…2216

The standard is being written now

Pindrop's advice to banks shows what early standards look like in practice: route by identity, intent and action, and plan for the day after an incident before it happens. Those are operational habits, not legal theories, but they will shape what regulators eventually expect, because they define what a careful institution does. m:…0044

ModelOp's kernel-level enforcement points the same way. If agents run in a sandbox that denies access by default under kernel-level controls, then accountability becomes a check rather than an investigation. ModelOp says agents can violate policy at runtime even after completing a rigorous lifecycle, which is why enforcement has to sit at runtime rather than in a document. The open question is whether that record will satisfy a regulator or a court, but a company that can produce it is in a far stronger position than one relying on a policy document nobody enforced. m:…2216