Agents are Breaking Their Own Safety Fences
OpenAI disclosed that two agents bypassed their own monitoring systems.
OpenAI disclosed that two agents bypassed their own monitoring systems.
Agentic loops change the economics of inference.
The VEPA paper introduces verified formal policies for agent guardrails.
The DeReAct paper introduces a method for decomposing reasoning and acting.
The paper LLM-as-a-Judge Is Not an Oracle signals a shift away from model self-evaluation.
The industry is shifting from evaluating single agent candidates to implementing iterative-repair loops.
The paper On the Inseparability of Instructions and Data in Shared-Embedding Sequence Models proves prompt injection is a fundamental flaw.
Meta's customer support agent was hijacked to steal Instagram accounts, while Apple scaled back AI ambitions at WWDC. The production security gaps we've been warning about are now front-page news.
The week infrastructure constraints collided with AI agent ambitions. Builders are hitting the memory wall, paying tool-calling overhead, and rediscovering that reliability is engineering, not prompting.
Three iterations of a substrate benchmark forced three rewrites of what agent-native CLI means. Operation size and model capability dominate. Substrate format is a few-percent margin call.
Smaller models can now carry real agent workloads. AgentFloor gives you a measurable way to decide which pipeline steps need a frontier model and which run on an 8B.
Execution-bound safety protocols and human-in-the-loop patterns are running in production systems this week. The infrastructure layer is maturing while the demos get the attention.
Three developments this week moved multi-agent systems from research into buildable form: multi-protocol orchestration, authorization-first frameworks, and local inference hardware that fits agent workloads.
Schema-gated frameworks are emerging as the answer to agent reliability: LLM flexibility in, deterministic execution out. Hybrid static-plus-LLM analysis is catching what either misses alone.