New Technique Medium TD-2026-0127 For AI Eng, AppSec, Red Team AI-assisted
Kaer: security-aware replanning to block feedback-based hijacking of multi-agent orchestrators
The paper describes Kaer, a multi-agent framework where the orchestrator's replanning is shielded from untrusted agent feedback. Diagnostics reach the planner only through abstract templates, and plan revisions are limited to deterministic repair operators that must satisfy preset policies. The authors evaluate Kaer on their Cintra benchmark against Magentic-One and ControlValve.
Why it matters
Multi-agent systems that revise plans from agent feedback can be hijacked by malicious agents embedding instructions in that feedback. Existing defenses that cover per-tool access or pre-execution policy leave this replanning phase unprotected.
Do this
- Treat inter-agent feedback as untrusted input and avoid passing raw agent text directly into orchestrator planning prompts.
- Constrain orchestrator plan revisions to a fixed set of approved repair operations and validate every revised plan against policy before execution.
- Log and review plan revisions triggered by agent feedback to detect unexpected changes in task scope.
- Evaluate candidate defenses against replanning-phase adversarial tests before deploying multi-agent workflows.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“This dynamic replanning introduces a critical vulnerability: malicious agents can embed instructions within their feedback to hijack the replanning process.”
“Kaer shields the planner from untrusted agent feedback by conveying diagnostics to the replanning procedure strictly through abstract templates.”
Sources
Affected / related
security-aware replanning, Magentic-One, ControlValve
Published 2026-11-18 18:05 UTC · updated 2026-11-19 04:35 UTC · AI-assisted summary, checked against quoted sources · How we score