Pre-AI, when Tim didn't know, the exposure was one question with no context. Now it's a standing connection to your repo and a retention layer you didn't configure.
Late Fee Media is back..
What changed: Tim, their engineer, used to ask Stack Overflow. He'd paste a snippet, keep it short, and move on. The exposure was a fragment that is frozen in time… Most important – he could delete it later and nobody was crawling those threads back then. Because who had the processing power and/or desire to process such data into usable analysis? Now? "Stack Overflow" sits inside the IDE with read and edit access to the repo, connections to internal systems and databases, and a context window that persists between sessions, which means the context is saved in another company's infrastructure. Last week Tim was debugging a playback failure he couldn'treproduce locally, so he pulled a few hundred rows of real viewing history in to the connected tool to work against actual data. Tim didn't break any policy or rule. But, all of a sudden Late Fee Media's customers' individual viewinghistory is now sitting with a vendor nobody diligenced, on a retention clock nobody set.
Why this matters: That window grants a use you never disclosed (001), on a retention clock you did not set (002), through a flow that is not on your map (003), to a sub-processor that is not in your inventory (004). Four wrinkles, one setting nobody checked. So, let's map it.
What your engineering and legal teams should be able to answer in the same room: • Licensing. Does the engineering team have enterprise seats - all of them, not just the ones who asked? (Enterprise seats + DPA reviewed by your legal?). • Training. Is training off at the organization level, and does an administrator gatekeep turning it back on? • Retention is a separate question from training. Turning training off does not turn logging off. Prompts, file contents, and saved context usually sit in their own layer with their own clock. How long does the vendor keep it, and does destruction align with the retention period in your own privacy notice? • Connectors. Who owns the final say on which connectors and APIs may reach which systems? And does the team know which categories of data never get connected at all? Ex: regulated personal data, health data, anything you hold under a customer's flow-down? • Security. Do the vendor's information security controls actually satisfy the standard you promised your own customers? You canonly flow down what your sub-processors hold. •Agent scope. What limits what an agent can reach when a task goes sideways - not just what stops an attacker? The likelier failure is broad credentials granted for a narrow task.
Pre-AI, when Tim didn't know, the exposure was one question with no context. Now it's astanding connection to your repo and a retention layer you didn't configure.
