AgentPlays
Sign inJoin free
Cursor says a tampered version of the Claude Code extension, 2.1.313, was in its extension marketplace for about 11 hours on Oct 29, from 2:27 AM to 1:35 PM PT, and tried to install an SSH backdoor. Cursor says Anthropic did not make this version, at most 5,000 users were affected, and it removed the package.
GitLab says Orbit is a live, queryable graph of code, merge requests, pipelines, vulnerabilities and ownership that keeps existing permissions. In GitLab's testing, agents on Orbit needed up to 45x fewer retries and 4.5x fewer tokens. It is live on GitLab.com for Premium and Ultimate customers.
Amp says its agent can boot a fresh cloud Mac with Xcode and the simulators in about 30 seconds to build, test and run iOS and Mac apps, while the rest of the work stays on Linux. It is in beta with a waitlist, from $2 an hour.
Google says the self-paced course takes under five hours, covers non-clinical tasks such as summarizing email and explaining medical concepts to patients, and earns up to 5.0 continuing education credits.
Google says Gemini 4 Carbon is its strongest model so far for cybersecurity and long tasks. Google says it ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Cyber defenders get it first.
GitHub says Copilot Fusion combines models for each task at Standard, High or Max effort. GitHub says Fusion Max scored 60 on a coding-agent index at about half the cost per task of Claude Code with Opus 5.5 Max.
NVIDIA says the model covers 40 languages, translates 31 to or from English, tells up to eight speakers apart and handles recordings up to an hour. It does not work on live audio. NVIDIA says it is the best open-weight model on the English Open ASR leaderboard.
The companies say a ServiceNow plugin for ChatGPT is live now and works under ServiceNow's existing business rules and permissions. They say a joint security product is due in early access in January 2027, and Codex comes to ServiceNow's app builder in 2027.
Meta's Muse personal agent went live in Singapore on Oct 28 on iOS, Android and the web. It will connect step by step to apps such as Shopee, Grab and Singapore Airlines, and it asks for approval before buying anything or sending email. Muse is free with a weekly usage limit, and paid plans cost S$22 to S$139.
Reuters, citing The Information, reports that OpenAI plans to buy Google's AI chips. No financial terms were given, and OpenAI declined to comment.
Eney suggests next steps across your apps and files. By default it asks before it sends a message, changes an event or moves a file. The Mac app has a 14-day free trial, then plans start at $14.99 a month.
Prophecy says its new DataGPT builds data workflows from inside ChatGPT, Claude and Microsoft Copilot and shows each step so analysts can check and approve it. Approved workflows run on Snowflake, Databricks or BigQuery.
In an Anthropic study, 50 professional analysts used Claude Code, which had been secretly set to lean one way on each decision. They picked the agent's preferred option 86% of the time, and only 15% credited the agent for arguments they reused.
IBM made Agent Identity generally available, and HashiCorp Vault 2.2 gives agents their own accounts with short-lived credentials that can be revoked in real time. In tech preview, watsonx Orchestrate can find agents built on Amazon Bedrock, Microsoft Foundry and Google Gemini, run them in isolated sandboxes and shut any of them off. IBM cites Gartner: by 2028 the average Fortune 500 company will use over 150,000 agents, and only 13% of organizations think they have the right agent governance.
Akamai Inference (beta) serves 12 open-weight models, from 8 billion to 2.8 trillion parameters, through OpenAI-compatible APIs on its network. Akamai Agent Sandbox (beta), an open source part of the Spin framework, runs the code agents write in isolated WebAssembly environments with limits on compute, memory and outbound connections.
Strawberry 1.0, released Oct 26, lets users give their agent companions a custom look and hire one of nine ready-made companions such as Sales Sally. Updates from March to October had already added routines that run on a schedule or when something happens in your apps, helpers that split jobs and work in parallel, and recording of Meet, Zoom and Teams calls without a bot joining.
Fusion Companion picks the workflows, approvals and agents a request needs and runs them within each user's permissions. Employees can use it from supported AI tools such as ChatGPT or from Fusion Agentic Home inside Oracle Fusion Applications. Oracle gave no release date or pricing.