The paper investigates failures in large language model (LLM) safety systems where runtime monitors detect policy‑relevant evidence but the evidence never reaches the enforcement decision point. Using canary contracts, the authors identified 18 evidence‑path failures across 11 safety projects and four ecosystems, categorizing them into established‑boundary repairs, new‑surface policy gaps, and boundary‑only evidence cases. The study demonstrates that ordinary application behavior can continue while safety‑specific evidence is lost, highlighting the need for regression tests on evolving monitor schemas, streams, tools, and adapters.
arXiv:2607. 19449v1 Announce Type: cross Abstract: Evaluation frameworks for tool-augmented LLM agents focus overwhelmingly on capability metrics or explicit tool crashes, leaving silent infrastructure failures and HTTP 200 responses with empty, null, or malformed payloads largely unaudited.
VEPA is a framework that enables the verified evolution of formal guardrails for lifelong large language model agents. It represents guardrails as formal policies of symbolic rules over typed agent events, and during deployment uses LLM‑guided feedback to propose policy patches that fix new failures while preserving correct behavior. A deterministic verifier accepts only structurally valid, conflict‑free policies, and an evidence‑aware authority layer activates verified rules only when sufficient supporting evidence exists, ensuring that uncertain updates do not directly affect guardrail decisions.
The paper introduces the Alignment Flywheel, a governance‑centric hybrid multi‑agent system (MAS) that separates decision generation from safety governance. It defines a Proposer that generates candidate trajectories, a Safety Oracle stack that evaluates safety, and an Enforcement layer that applies risk policies at runtime. A governance MAS oversees monitoring, red‑teaming, verification, and versioned release management, enabling patch‑local fixes to safety failures without retraining the Proposer. The architecture is implementation‑agnostic and is demonstrated in two scenarios: a learned spatial Oracle and a clinical GenAI proxy. The authors provide open‑source code at https://github.com/decide-ugent/Alignment-Flywheel.
The paper introduces Runtime Assurance Contracts (RAC) as a formal policy framework for high‑risk AI agents, addressing the "assurance‑transition gap" by binding autonomy boundaries, component eligibility, evidence state, transition policy, human‑review capacity, and non‑compensatory gates. RAC allows soft metrics to influence routing while mandating retries, switches, escalations, deferrals, or stops when mandatory gates fail or are unknown, ensuring aggregate performance cannot alone authorize action. The authors define the contract, evidence record, permission rule, and five invariants, and evaluate RAC through deterministic failure‑injection studies, hand‑authored traces, and a prospective synthetic holdout, comparing it to score‑only and restricted protocol baselines.
arXiv:2608. 12654v1 Announce Type: new Abstract: Long-running LLM agents act through tools, and a single step can send an email, merge a pull request, or wire a payment.
arXiv:2608. 11274v1 Announce Type: cross Abstract: The dominant paradigm treats AI safety as a property to be instilled during model training via RLHF, DPO, or Constitutional AI.
arXiv:2607. 01793v1 Announce Type: new Abstract: LLM agents increasingly perform autonomous actions through external tools, leading to complex and evolving safety risks.
arXiv:2607. 12469v1 Announce Type: cross Abstract: Many agent-safety evaluation results are not yet load-bearing evidence: identical nominal outcomes (task success, attack success, monitor scores) may sit atop materially different evidence regimes.
arXiv:2607. 07097v1 Announce Type: new Abstract: Safety evaluations of multi-agent LLM systems often compare a direct prompt with a planner-executor pipeline and report the difference as a single "pipeline effect.
arXiv:2605. 27784v2 Announce Type: replace Abstract: LLM agents are governed by long-lived prompt policies, where individually reasonable stand- ing rules can jointly govern the same pre- generation state.
The paper introduces the twin‑prefix framework to evaluate how the size of the verification unit—i.e., how many actions a pre‑execution LLM monitor reviews in one call—affects its performance. By pairing each gold plan with a twin that differs by a single write and injecting a controlled error, the authors isolate the impact of review length on catch rates and false rejections. Their findings show that longer review windows increase rejection rates but do not improve discrimination, with the highest informedness occurring at one or two actions across all judges and domains.
arXiv:2606. 04990v1 Announce Type: cross Abstract: Large language model (LLM)-based agents increasingly solve complex tasks by interacting with external tools, retrieval systems, memory modules, environments, and other agents.