Tuesday 15 September 2026Paper No. 1,386

Evening EditionRSS

ZOTPAPER

News without the noise

Safety & Evals

Guardrails, attacks, red teaming, benchmarks

44 articles

All Agent Harnesses Tools & MCP Multi-Agent Systems Creative Agents Memory & Planning Safety & Evals RL for Agents AI & ML

Viewgrid list compact

Hidden prompt revisions inject cultural stereotypes into generated images

Deep research agents struggle with long, multimodal evidence chains

Refresh schedules, not cache age, determine agent answer staleness

Cross-device workflows expose major weaknesses in today’s GUI agents

API benchmark scores poorly predict performance in consumer chatbot interfaces

Security contracts preserve authorization across composable LLM agent systems

Current robot models falter as spatial and procedural complexity rises

Causal framework distinguishes deceptive outputs from deceptive mechanisms in language models

LLM judges as measurement instruments fail reliability tests on shared endpoints

Graph-based policy constrains LLM agents for topology-aware incident response

Coding agent benchmarks miss review compliance gaps

Digital twins make agent decisions replayable, but hidden tool states remain

Malicious plugin updates can hijack AI agent lifecycle hooks

Post-training choices reshape refusal circuits but leave safety trade-offs

Deterministic guardrails expose gains that fool LLM-based judges

Structured behavioral abstractions improve diagnosis of failures in LLM agents

Adaptive logit tilting exposes rare language-model behaviors more reliably

LLMs fail at multi-round code review in new realistic benchmark

Process-level tests expose distinct mathematical agent capabilities in LLMs

Separating tool suggestions from authorization sharply reduces agent attacks

Polished evidence makes LLM agents act on unknowable questions

MLLMs fail to sustain goal-directed navigation in real-scale 3D city

LLM agents struggle to infer wellbeing from long-term wearable data

Agent safety requires persistent state across autonomous loop iterations

Domain context improves LLM agents’ time-series root-cause attribution

Plan-first controls block attacks on persistent language-model agents

Dedicated intent tools expose shifts toward harmful agent behavior

Tracing and fault injection make multi-agent software workflows easier to inspect

Reinforcement learning strengthens policy invocation for agent safety judgments

Semantic scores overstate language-model safety in air traffic control

Bayesian framework exposes hidden differences among seemingly similar RAG systems

Medical AI models often give right answers with disconnected reasoning

In AI-Q, most report errors originate in the orchestrator

Adaptive trace graphs improve failure attribution in multi-agent systems

Physiological grounding makes clinical language-model recommendations safer in controlled tests

Step-level checks curb unsafe agent actions with little utility loss

Financial agents can cite rules while still attempting prohibited trades

Probabilistic scoring unlocks verification as a new scaling axis for LLMs

Masking untrusted web content to secure agents against prompt injection

Persistent memory in AI agents can be stealthily poisoned via a single email

Persistent-codebase AI agents vulnerable to distributed, gradual attacks

Automated safety testing reveals 93.9% attack success rate across four agent frameworks

First measurement study reveals widespread authentication flaws in remote MCP servers

Benchmark reveals frontier LLM monitors miss most covert attacks by coding agents