Browsing: AI agents
Reporting on AI agents that browse, code and act on your behalf, covering what they can really do today and where they still go badly wrong so far.
OpenAI disclosed six new cases of concerning model behavior, including models leaving secret notes telling future versions of themselves to hide mistakes. Here is what the six cases show and what they don’t.
Cursor’s new Projects beta hands a coordinator agent the job of managing thousands of subagents at once. Early users are merging far more code. They are also learning what that costs.
GreyNoise says a single attacker ran hundreds of AI agents combining OpenAI’s Codex and DeepSeek to breach 440 PaperCut servers across 395 organizations in 48 countries, reaching remote code execution in under four hours.
OpenAI opened its Agents API to public beta on September 10, exposing the Codex harness behind one API call. Early testers report 86% fewer failed responses, but the orchestration layer now belongs to OpenAI.
OpenAI says it has reached its automated research intern milestone, with agents now producing 3.1 agent workdays of effort per human workday. Every figure in the claim comes from OpenAI.
Instinct’s AI assistant sent an email without approval and kept summarizing an inbox after access was revoked. Days later, investors valued it at $2.5 billion.
Nvidia’s AVO agent harness scored a perfect 100% on the ARC-AGI-3 benchmark using the same Claude Opus 5 model that scores just 30% on its own, without changing a single model weight.
Nvidia’s Open Secure AI Alliance grew from 37 to more than 120 members in two weeks and published its first SAFE security guidelines at Black Hat. OpenAI, Anthropic and Google still are not in it.
OpenAI’s new Apple Messages plugin lets ChatGPT read, draft and send texts from your Mac. In its own documentation, OpenAI warns against persistent send approval. Here is what the plugin does, why prompt injection makes that warning serious, and how to switch the setting back.
SpaceXAI’s Grok 4.6 chases GPT-5.6 Sol with big coding and long-agent benchmark gains, ships to Cursor and OpenRouter on day one, but skips a safety card for the second release in a row.