If your agent releases or trades on material non-public information (MNPI), you are just as liable as if you released the information or made the trade yourself. Trustabl.ai provide your with the tools to prevent leakage of MNPI as well as the use of MNPI.
Post 5 of a series on AI liability for general counsels interested in the legal liability associated with building and using agents.
Your agent just pulled last quarter’s numbers, cross-referenced them with a customer win that has not been announced, and drafted a note for the sales team. Or it answered a question from an analyst call transcript with a detail that is not yet public. Or it updated a forecast model with data that only three people inside the company are supposed to have. None of those steps required the agent to “decide” to trade. But each one created potentially costly exposure.
Material non-public information (MNPI) does not care whether the entity that exposed it was a person or an agent. Under Reg FD and the insider-trading rules, the firm must control who has access and when information is disclosed. An agent that can read from systems containing MNPI, write to systems that leave the firm, or hand context to another agent creates the same problem a careless employee creates, only faster and with a weaker record of who authorized the path.
Most companies treat this as a data-access issue. The agent can read the sensitive information. It can also call a tool that posts to Slack or drafts an email. The policy that was supposed to stop the combination lives in one place. The identity the agent represented lives in another. When the question arrives (from compliance, from a regulator, or from opposing counsel) the answer is a reconstruction across systems.
Composition makes the problem worse. A research agent with read access and a communications agent with write access can, through handoff or shared context, produce a disclosure neither was authorized to make alone. An agent that is allowed to summarize internal documents and an agent that is allowed to answer external questions can meet in the middle. The individual ceilings were respected. The combined outcome was not.
The company’s agents didn’t attend the internal training on MNPI. Even if you trained them on the subject, like all guardrails, they take it as a suggestion, not enforced policy. You need data-class restrictions that treat MNPI as a distinct class. You need explicit rules on what may be read, what may be written, and what may leave the agent. You need identity binding so you can show which human or workload the agent represented. You need human approval on actions that could constitute disclosure. You need the intersection of those ceilings compiled into a signed contract the runtime enforces. You need traces that record which data classes were touched, which tools were called, and which policy version was in force. Unfortunately for you, those tools are typically separate and they don’t collaborate on policies that are actively enforced to prevent the disclosure.
At Trustabl we treat MNPI exposure as a data-class and handoff problem rather than a separate product. Company Policy can declare material non-public information as a restricted class, set read versus write permissions, and prohibit copy-out or external communication without approval. The compiler intersects those ceilings with the other constraints and produces one signed contract. It generates the corresponding policies for the Policy Enforcement Points (PEPs) so the restriction holds across agents instead of stopping at the first one. The Safe Agent Control Plane records the data classes touched, the handoffs, and the approvals in the same evidence bundle. We do not claim this eliminates every disclosure path. We claim it turns an invisible combination of read and write rights into something you can see, constrain, and prove you constrained.
Agents will keep being asked to summarize, forecast, and communicate. The companies that can show the agent was not permitted to turn internal data into an external disclosure will be in a different position from the companies that discover the path after the information has already left.
Additional reading in this series
Related Blogs
Agent Reliability Resources
