Kaer Framework Beats Rivals on Security Replanning
Researchers Tushin Mallick, Evan Rose, Matthew Laws, Alina Oprea, and Cristina Nita-Rotaru published a paper on November 18, 2026, introducing Kaer, a multi-agent framework built for security-aware replanning, with the arXiv record updated November 19, 2026 (arXiv).
The Problem: Hijacked Replanning
The paper focuses on multi-agent systems in which an orchestrator generates, executes, and revises plans based on feedback from the agents it coordinates (arXiv). The authors identify a specific vulnerability in this setup: malicious agents participating in the system can insert instructions into their feedback, attempting to hijack the orchestrator's replanning process and steer it toward attacker-chosen outcomes rather than the original task objective.
This is a distinct threat model from simple prompt injection against a single agent — it targets the control loop that lets multi-agent systems adapt when something goes wrong, turning the system's own resilience mechanism into an attack surface.
How Kaer Defends Against It
Kaer's defense rests on three mechanisms described in the paper. First, diagnostics from agent feedback are passed into the replanning process only through abstract templates, rather than as raw, freeform text that could carry embedded instructions. Second, the framework limits any plan revision to a fixed set of deterministic repair operators, constraining how much a single piece of feedback can alter the overall plan. Third, every modified plan must satisfy preset policies before it can be executed, acting as a final check against hijacked outputs (arXiv).
Benchmarking on Cintra
To evaluate the approach, the authors also introduce Cintra, a benchmark built specifically for tasks involving replanning under adversarial objectives — scenarios designed to test whether a multi-agent system can be manipulated through its own feedback loop.
Against this benchmark, Kaer was compared to two systems: Magentic-One, described in the paper as an undefended multi-agent system, and ControlValve, described as the current state-of-the-art defense. The reported result is that Kaer achieved greater robustness than both comparators, while maintaining utility — meaning task performance on legitimate objectives — comparable to ControlValve (arXiv).
Why It Matters
As orchestrator-style multi-agent systems move from research demos toward production deployments, the replanning loop becomes a practical attack surface wherever one agent's output feeds into another's control decisions. Kaer's template-based diagnostics and policy-gated repairs offer one concrete architectural pattern for closing that gap without discarding the adaptability that makes replanning useful in the first place. The paper does not report any court, regulatory, or commercial action tied to the findings — the contribution is research-stage, published via arXiv.


