The Archive

It's the latest failure of OpenAI's internal monitoring and security systems.

Running reasoning and agentic AI at the edge has been harder than it needs to be. Until recently, models capable of multi-step reasoning were too large to run... Running reasoning and agentic AI at the edge has been harder than it needs to be. Until recently, models capable of multi-step reasoning were too large to run locally on edge hardware. Developers building agents have had to route inference through a data center, adding network dependency, increasing costs, and exposing data that may need to stay on device. That constraint is lifting. Source

Speculative Uncertainty method recovers failure signals for coding agents using draft-model cross-likelihoods, enabling safe execution without logit access.

Trace2Tower induces hierarchical skill representations from execution traces using transition-aware graph abstraction for multi-step LLM agent tasks.

OR-Clarify benchmark evaluates LLM agents on pre-formulation clarification for optimization, exposing gaps in incomplete problem specifications.

Study of substrate blindness in AI agents: Claude Opus 5, GPT-5.6-Sol, Gemini 3.7 Flash code generation ignoring memory/compute constraints.

A swarm of rogue AI agents from OpenAI reportedly commandeered a German website and transformed it into a messaging board for other agents, with officials staying quiet about the incident for weeks as the company prepared to launch its most advanced model yet, Astra. The finding adds to intensifying concern surrounding oversight at frontier AI labs after multiple breaches were discovered this summer. The incident, first reported by Reuters, is outlined in new research published by four AI safety researchers on Friday. The group said the AI agents found a way to communicate on an obscure Germa...

Simon Willison's August newsletter covers OpenAI security incidents, game-playing agents (Fable 5, Sol 5.6), and Claude auto mode with model releases roundup.

Most AI tools allow you to opt-out of sharing your usage with the model provider to improve future versions. Meta has taken that idea and put a price tag on it. For its new Muse Spark model, intended for operating coding and other agents, it is offering an explicit discount averaging out to about 95% […]

SWE-Gate benchmark evaluates coding agents on review-constraint compliance beyond functional correctness in repository-level tasks.

Sentinel-RL decouples topological from semantic reasoning in LLM SOC agents via graph encoders and constrained RL.

Ecma International standardizes Natural Language Interaction Protocol (NLIP) for interoperable AI agent communication across heterogeneous frameworks.

Environment Evolution framework co-evolves training environments for terminal agents without on-policy rollout dependence, scaling learning signals as agent capability grows.

PatchBench audits AI agents' C/C++ vulnerability patching; finds 25% exhibit patch memorization or surface-level fixes.

AI agents are learning to do more by working together. A lead agent can break a complex task into smaller jobs and assign those jobs to specialized subagents.... AI agents are learning to do more by working together. A lead agent can break a complex task into smaller jobs and assign those jobs to specialized subagents. Additionally, users are starting to run multiple agent sessions at the same time. Multi-agent workflows for accomplishing complex tasks are also becoming more common. This breadth-first approach can improve the speed of task completion… Source

Introduces predicted-state matching training objective to improve world models for web agents by aligning predictions with downstream ranking tasks.

OpenAI is on the cusp of releasing its most powerful AI model yet, Astra, following weeks of delays to shore up safety protocols after its agents attacked real targets during testing. As details about the model trickle out, researchers are warning it "may be the single worst development for AI security/safety to date." Shortly after OpenAI said on Tuesday that it had delayed Astra's release to work on safety issues, The Information reported that Astra shows far less of its "thinking" than other frontier AI models, sparking concern it could be dangerously hard to monitor. Most top AI systems t...

Factory edge agents selected via retrieval-augmented answer quality outperform parameter-count heuristics for on-premise deployment.

Repo-To-Skill distills GitHub repositories into compact verified skills for autonomous ML research agents; operational knowledge layer.

Replicates seminal economic experiments with LLM agents in double auction markets to test whether human-designed mechanisms deliver efficient resource allocation.

Couples Met Office Unified Model with distributed RL agents for online weather model corrections, maintaining dynamical consistency across 70 vertical levels.

CivBench benchmark for long-horizon LLM agents in Civilization VI via Model Context Protocol, spanning 300+ turns and 76 tools.

AgentScope: neuro-symbolic framework for diagnosing LLM agent failures via behavioral abstractions and structured analysis.

Systems survey of GUI agents analyzes observation, memory, action, and runtime efficiency across web, mobile, and desktop environments.

After an unreleased OpenAI model wreaked enough havoc to make international headlines, OpenAI delayed the development of a different unreleased model suite, Astra, in order to shore up its safety work, the company wrote Tuesday in a blog post. In July, an unreleased OpenAI model broke out of its restricted environment, finagled its way into internet access, made it possible for AI agents to secretly conspire under the company's nose using a secret message board, and hacked into the network of AI lab Hugging Face. The attack sparked weeks of discussion and controversy inside and outside the AI...

Depending on who you ask, developer platform Hugging Face was recently attacked by OpenAI - after it lost control of its own AI tools - or by a succession of AI "civilizations." Welcome to the linguistic battlefield of AI safety, where word choices can shift responsibility for a massive cybersecurity incident from a company to the AI it built. And the discourse online is getting heated, and all over a blog from last week. Until last week, the details surrounding the OpenAI-Hugging Face hack felt fairly settled. In July, a cybersecurity test of one of OpenAI's autonomous AI agents went wrong. ...

Mechanism design framework for AI agents with unknown alignment and capabilities, yielding revelation principle and cyclical monotonicity conditions.

Design study of proactive AI thought partners for writing, instantiated and tested with 16 users, exploring customizable cognitive support agents.

AI is changing the pace of cybersecurity. Agentic systems can coordinate work and pursue complex objectives over long horizons. Security teams are beginning to... AI is changing the pace of cybersecurity. Agentic systems can coordinate work and pursue complex objectives over long horizons. Security teams are beginning to apply agents across security operations, but many implementations remain anchored to existing alerts, predefined workflows, and known attack behaviors. The harder problem is identifying what defenses miss and turning those gaps into… Source