Agentic AI

Researchers introduce masked diffusion language models as a solution to the left-to-right bias inherent in autoregressive world models. This approach enables better conditioning on globally interdependent state anchors like tool schemas and expected outcomes. The result is a steerable text-based world model that supports diverse, on-demand training environments for reinforcement learning agents.

• Addresses mode collapse in RL caused by sparse rewards and fixed task difficulties.
• Overcomes autoregressive limitations by conditioning on global state anchors.
• Enables on-demand diversity scaling for specialized agentic training environments.
• Formalizes text-based world modeling as a steerable transition dynamic.

Researchers identify the planning phase in multi-agent LLM systems as a critical vulnerability where a single prompt injection can cascade through downstream executors. The PlanFlip framework demonstrates four attack vectors—GoalSubstitution, PriorityInversion, ContextPollution, and RoleConfusion—that mimic legitimate tool outputs to bypass filters. Testing across nine frontier models reveals that higher capability often correlates with increased susceptibility to these specific planning-phase attacks.

• Multi-agent architectures are vulnerable at the Planner stage, not just execution.
• Attacks disguise malicious prompts as plausible tool outputs to evade keyword filters.
• Higher model capability does not guarantee immunity; it may increase risk.
• Four distinct attack types (PF-1 to PF-4) systematically corrupt sub-task sequences.
• Defenders must inspect Planner context integrity, not just final outputs.

AI agent systems combining LLMs with external tools suffer from inherent non-determinism due to sampling variance and external state. Existing observability tools capture logs but cannot reproduce these runs in isolation. The authors introduce agrepl, a CLI framework that uses a MITM proxy to intercept and serialize external interactions at the transport layer. This enables strict replay in an isolated environment with zero outbound network access.

• Solves the non-determinism problem inherent in LLM-agent interactions with external APIs.
• Uses a MITM proxy to serialize external interactions into structured execution traces.
• Enables faithful reproduction of agent runs in an isolated, zero-network environment.
• Provides a developer-first CLI for debugging and testing agent behavior deterministically.

MOSAIC addresses two critical failures in current long-term memory systems: the loss of relational context in flat storage and the high latency of LLM-based classification. By implementing an entity-typed graph structure, it preserves the temporal and multi-hop relationships necessary for accurate reasoning. The framework also introduces conflict-aware mechanisms to validate new data against existing knowledge, preventing the silent accumulation of contradictions.

• Replaces flat storage with entity-typed graphs to retain relational context for complex queries.
• Eliminates expensive LLM-based classification steps to reduce latency for real-time agents.
• Detects and resolves contradictions between new inputs and stored knowledge to maintain accuracy.
• Enables reliable multi-hop and temporal reasoning without hallucinated or lost context.