Research
Long-running AI agents quietly drop the compliance rules they started with, and a bigger context window doesn't fix it
New reporting on 'context poisoning' finds that as an agent's session grows, it gradually deprioritizes the compliance directives set at the start, with the Cloud Security Alliance putting the number of organizations reporting agents exceeding their intended permissions at 53%.
The mechanism is specific enough to matter for anyone running an agent on a multi-day workflow rather than a single-shot query: as tool calls and retrieved data accumulate in context, that material can subtly contradict or crowd out the compliance directive that was set in the system prompt hours or days earlier. VentureBeat's reporting frames this as attention dilution, not a bug in any one model, a million-token context window still suffers from it, because the problem is which instructions the model weights, not how many tokens it can technically hold.
This is the same fault line this site has been tracking since the CFTC's Innovation Advisory Committee agenda and the FINRA Rule 3110 human-in-the-loop requirement: both assume a compliance rule set at deployment time stays enforced for the life of the session. An autonomous trading agent running for days against live market data, adjusting positions, re-reading its own prior fills, is exactly the long-running, tool-heavy workflow this research says is most exposed to the drift. A firm that satisfied its supervisory obligation by hardcoding a risk limit into a prompt at session start has no guarantee that limit is still load-bearing by day three.
The recommended fix, moving hard constraints out of the LLM's context and into a deterministic policy engine that can veto or block an action regardless of what the model currently believes its own rules are, is a more expensive answer than prompt engineering, and it is also the only one that doesn't degrade with session length. For any desk treating a system prompt as its actual control, rather than a wrapper's suggestion sitting in front of an independent enforcement layer, this is a reason to check which one they actually built.
Read the original: VentureBeat - Long-running AI agents quietly drop the compliance rules they started with, and a bigger context window doesn't fix it. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.