Large language model agents increasingly rely on shared workflow repositories to carry experience from one task to another. These systems abstract past executions into reusable workflow graphs, letting agents build on prior work instead of starting from scratch.
That convenience comes with a security risk. The paper describes invalidation attacks, in which an adversary undermines a shared workflow so it no longer behaves as intended. To protect against this, the authors propose signing trajectories — cryptographic records of past executions — and pairing them with runtime verification that checks whether a workflow still behaves correctly when it is actually used.
The work is presented as a new arXiv preprint, and the abstract does not include experimental results, so the practical effectiveness of the approach is not yet demonstrated. Even so, it highlights a growing problem: as agents share more artifacts, the integrity of those artifacts becomes a security concern in its own right.