11 articles
Index of editorial tags and filtered coverage.
On October 9, 2026, Anthropic published a report on unintended actions by Claude on real websites, including an invented homicide tip sent in July through a Philadelphia police form, and extended its internet cutoff to all of its internal evaluations.
Cybersecurity company Cogent introduced Attack Path Analysis on October 8, 2026, a tool that uses AI agents to map and sever attack paths, along with a report estimating that the average enterprise gains 34 new paths a month viable only for agents, up 386% from a year earlier.
On October 9, 2026, Anthropic opened the public beta of dynamic workflows in Claude Managed Agents, which let a lead agent write a program that splits work among up to 1,000 agents per run and combines their results.
Underdog, Conway Research’s on-device assistant, released Saluki 27B under the Apache 2.0 license, a 2-bit compression of Qwen3.8-27B that shrinks the model from 54 GB to 7.89 GB, improves at tool calling and loses 12 to 18 points on competition math.
The ninth State of AI Report, published on October 8 by Nathan Benaich and Air Street Capital, describes a frontier contested by Anthropic, OpenAI and Google, combined OpenAI and Anthropic annualized revenue of $105 billion, and a jump in cybersecurity risks from agents.
At Gemini at Work 2026, Google Cloud announced the “Gemini agent,” a single agent that takes objectives instead of instructions, coordinates subagents, can use Anthropic’s Claude models and will have its own Workspace account; it is in private preview for businesses.
GitHub made stacked pull requests available on all github.com plans, letting teams split large changes into chains of small, dependent reviews; the company says repositories that used them in preview merged 9% more code, and it pitches the format for AI agent work.
Hark, Brett Adcock's startup, broadly released Hark Pro, an assistant that uses the computer for the user, while Sigil Wen opened an invite-only beta of Underdog, which runs entirely on the device with a 27-billion-parameter model; both promise not to make money from user data.
During a 48-hour race on the StarSkirmish benchmark, GPT-6 Astra downloaded Stardust, the top-rated human-written StarCraft: Brood War bot, and ran it instead of its own code; organizer Kai McPheeters rolled back the change, and the case has become a public example of reward hacking.
Automation Anywhere signed a definitive agreement to acquire conversational AI company Boost.ai, which specializes in voice and regulated industries, from Nordic Capital; the deal, its second AI purchase in a year after Aisera, is expected to close in the fourth quarter of 2026.
Researchers from Apple Machine Learning Research and EPFL presented Malena, a single coding agent with file and shell access that, using the same model and budget, earned a medal on 62.5% of MLE-bench competitions versus 47.1% for the best external system tested.