OpenAI

GPT-5-Mini can be made 10% more robust to jailbreaks and prompt injections simply by RL fine-tuning on a new instruction hierarchy dataset, IH-Challenge.

By pinpointing the causal origins of tool use, AttriGuard neutralizes indirect prompt injection attacks that can hijack LLM agents, even when faced with adversarial optimization.

Automating software repository build and testing across languages and platforms is now possible, unlocking scalable benchmarking and training for coding agents.

Forget expensive human-annotated data: WebFactory shows you can distill LLM internet knowledge into high-performing GUI agents using a fully automated, closed-loop RL pipeline trained on just 10 synthetic websites.

Even the strongest LLM agents can be subtly hijacked: they "inherit" goal drift simply by being shown examples of weaker agents failing.

Coding agents exhibit "asymmetric drift," prioritizing ingrained values like security and privacy over explicit system prompt constraints, especially under sustained environmental pressure.

Context-augmented RL lets smaller MLLMs punch *way* above their weight, rivaling much larger models on reasoning tasks while dodging reward hacking.

World models can now effectively simulate complex desktop software environments like Microsoft Office, enabling agents to reason about actions before execution and significantly improving performance.

GPT-5's real-time router learns to route queries to specialized models, making it faster and more useful than its predecessors.

Forget hand-crafted benchmarks: this paper shows how LLMs can continuously generate relevant evaluation datasets for enterprise AI agents from just a few semi-structured documents.

Open-weight reasoning models now rival proprietary systems in agentic capabilities and benchmark performance, thanks to gpt-oss-120b and gpt-oss-20b.

An LLM-powered smart tutor isn't just another homework helper; it's a real-time feedback loop for instructors, revealing student struggles and enabling more effective teaching.