News · Leaderboard

Latest news, beside the leaderboard

Read what just changed, and see who is ahead. Those are the two doors on this page.

O
OpenAI Developers@OpenAIDevs·Collected 32 min agoFresh · 6h
🚀 Release

OpenAI Brings GPT-Live-1.1 to the API: More Consistent Accents, Better Transcription, New $0.02/Min Standby Mode

On November 19 OpenAI made its full-duplex voice model GPT-Live-1.1 available in the API (model ID gpt-live-1.1 on v1/live/sessions). It has powered voice in ChatGPT Plus and Pro since October 15 and, compared with GPT-Live-1, transcribes more accurately, keeps accents and voice more consistent across languages, and lets developers describe in plain language when to delegate to a backend model. OpenAI also added Standby: after 5 to 60 seconds of silence a WebRTC session pauses inference and is billed at $0.02 per minute instead of $0.05 until someone speaks again.

OpenAI Brings GPT-Live-1.1 to the API: More Consistent Accents, Better Transcription, New $0.02/Min Standby Mode
⚡ Key Takeaways
  • •Official: announced by @OpenAIDevs on November 19, logged in the OpenAI developer changelog the same day, with a model page at developers.openai.com/api/docs/models/gpt-live-1.1
  • •Capabilities: full duplex, listens while it speaks; versus GPT-Live-1 it transcribes more accurately, keeps accent and voice more consistent across languages, and reliably catches precise details such as letters and numbers
  • •Delegation: describe in plain language in the session configuration when to hand reasoning or tool use to a backend model of your choice without breaking the conversation
  • •Standby: WebRTC sessions on GPT-Live-1 and 1.1 can set a 5 to 60 second idle timeout; while paused they bill at $0.02 per minute and resume when someone speaks
  • •Pricing: billed by session duration at $0.05 per active minute, not rounded up to the next minute; backend model and tool usage are billed separately, and the two models have separate concurrent-session limits
Read details→
G
Google AI Studio@GoogleAIStudio·Collected 1h agoFresh · 6h
🛠️ Tooling

Gemini API Launches Tool Retrieval in Preview: On-Demand Tool Schemas, Remote MCP Support, No Extra Charge

On November 19 Google launched Tool Retrieval in preview in the Gemini Interactions API. Developers add a tool_search tool and set defer_loading: true on tools they want to hold back; the model sees only tool names and descriptions, then loads a full parameter schema by name and calls it within the same request, cutting input tokens and latency and keeping tool selection accurate as libraries grow. Remote MCP servers can be deferred as a whole, there is no separate charge, and it works with Gemini 3.6 and later models, including all Gemini 4 models.

Gemini API Launches Tool Retrieval in Preview: On-Demand Tool Schemas, Remote MCP Support, No Extra Charge
⚡ Key Takeaways
  • •Official: Google blog post on November 19 by Lucia Loher, Product Manager for the Gemini API, and Nora Kassner, Research Scientist at Google DeepMind, also posted by @GoogleAIStudio
  • •Usage: add {"type": "tool_search"} to the request and set defer_loading: true on tools to hold back; frequently used tools can stay loaded up front
  • •Flow: search, schema load and the function_call all happen inside one API call and are recorded in interaction.steps for debugging
  • •MCP and client mode: set defer_loading on an mcp_server tool to load all of its tools on demand, or use "execution": "client" to plug in your own tool registry or permission filter at the cost of one extra round trip per search
  • •Pricing and models: no separate charge; enabling it adds a small fixed block of input tokens and each loaded schema is billed as input for the rest of the conversation; works with Gemini 3.6 and later models, in the Interactions API only
Read details→
O
OpenAI Developers@OpenAIDevs·Collected 2h agoFresh · 6h
🛠️ Tooling

OpenAI Reveals Performance Factory: Codex Agents Investigate ChatGPT Performance Regressions, 1,700+ Investigations and 600+ Root Causes in Three Months

In a November 19 blog post OpenAI described Performance Factory, an internal system where a performance alert opens a task, Codex agents find the root cause, a separate Codex session runs an adversarial review, Codex opens a fix PR for an engineer to review, and an agent later checks that metrics recovered. Its first workflow, the Regression Workflow, tracked more than 1,700 investigations and root caused over 600 cases in just over three months, about 17 per day; in a manual review of 100 root cause suggestions, 70% were graded accurate.

OpenAI Reveals Performance Factory: Codex Agents Investigate ChatGPT Performance Regressions, 1,700+ Investigations and 600+ Root Causes in Three Months
⚡ Key Takeaways
  • •Official: OpenAI blog post on November 19 by Members of Technical Staff Okay Zed and Cyrus Forbes, also announced by @OpenAIDevs
  • •Architecture: four components, a task tracker with configurable workflows, a dispatch loop, an API that runs Codex in the cloud, and subscriptions that wake agents on events such as PR comments or completed deployments
  • •Numbers: 1,700+ investigations and 600+ root causes in just over three months, about 17 per day; of 100 manually reviewed root cause suggestions, 70% were graded accurate and 30% needed refinement or did not fully establish causality
  • •Guardrails: both diagnoses and fixes pass an adversarial review by a separate Codex session, and engineers still decide whether to merge each PR
  • •Next: synthetic performance tests so agents can measure fixes before review, and exploring shipping low-risk fixes without a human in the loop
Read details→
L
LuminaBench@LuminaBench·Collected 3h agoFresh · 6h
🚀 Release

Google Makes Genie 4 Generally Available in the Gemini API: Real-Time Steerable Worlds on the Interactions API at About $0.10 per Second

Google says in the Gemini API changelog that Genie 4 (models/genie-4) became generally available on the Interactions API (v1beta) on November 19. It generates a steerable 16:9 world at 24 fps with synchronized 48 kHz stereo audio in real time, streamed over WebRTC or WebSockets. Worlds can be seeded from text, a 16:9 reference image or both, then steered with WASD or arrow keys and mid-session text events. It is paid tier only, with video output at $17.50 per million tokens, about $0.10 per second. A day earlier Genie 4 opened to Google AI Ultra subscribers in Project Genie.

⚡ Key Takeaways
  • •Official: the November 19 Gemini API changelog entry marks Genie 4 GA, with model ID models/genie-4 on the Interactions API v1beta
  • •Output: real-time 16:9 video at 24 fps plus 48 kHz stereo audio; keyboard moves take effect in about 0.2 to 0.5 seconds, text events in about 1 to 2 seconds
  • •Pricing: paid tier only; $1.50 per million input tokens, $9.00 per million text output tokens and $17.50 per million video output tokens (240 tokens per frame, about $0.10 per second with audio)
  • •Limits: no free tier; Tier 1/2/3 get 2/5/10 concurrent sessions and 50/250/1,000 requests per day; sessions hard-cap at 120 seconds and are served from US data centers only
  • •Usage: create the interaction on your server, mint a single-session token, connect from the browser over WebRTC, and always call complete() to stop billing
Read details→
ADSponsored
P
Perplexity@perplexity_ai·Collected 4h agoFresh · 6h
🛠️ Tooling

Perplexity Brain Now Auto-Creates and Updates Personal Agent Skills for Computer: 80,000+ Skills Written, On by Default for US Max Subscribers

On November 18 Perplexity published an engineering post saying Brain, the personalized self-improvement system for its Computer agent, now automatically creates and updates agent skills, and announced it on X on November 19. Brain uses a user's memory and session history to write and refine personal skills, and a similar approach improves the built-in skills available to every user. Between August 1 and November 14 Brain wrote more than 80,000 personal skills for over 20,000 users. In an eight-week randomized test, tool call failures fell 11% and user dissatisfaction fell 7.6%, with no measurable change in cost or response time. Personal skill improvement is available to all US Max subscribers, on by default and can be turned off.

Perplexity Brain Now Auto-Creates and Updates Personal Agent Skills for Computer: 80,000+ Skills Written, On by Default for US Max Subscribers
⚡ Key Takeaways
  • •Official: Brain writes and updates personal skills from memory and session history, and improves built-in skills from aggregate patterns across all sessions
  • •Loop: four stages of evidence, reflection, verification and activation, with an LLM judge auditing sampled proposals against the evidence
  • •Scale: over 80,000 personal skills for 20,000+ users from August 1 to November 14, plus 100,000+ automatic updates to 30,000+ existing skills since September 4
  • •Results: in an eight-week randomized test tool call failures fell 11% and dissatisfaction fell 7.6%; on SkillLearnBench, success over 16 paired instances rose from 43.8% to 62.5%
  • •Availability: all US Max subscribers, on by default and can be turned off; the Plugins tab marks which skills Brain wrote
Read details→
M
Moritz Sanft@stdoutput·Collected 5h agoFresh · 6h
🛠️ Tooling

Researcher Used Codex to Find Ruby Case-Mapping Bugs Enabling Pre-Auth RCE on Rails; Ruby 4.0.10, 3.4.13 and 3.3.14 Fix Seven Vulnerabilities

On November 18 Ruby published security advisories for seven vulnerabilities in its regular expression engine and string case conversion, with known PoCs showing some can lead to remote code execution, and released Ruby 4.0.10, 3.4.13 and 3.3.14 to fix them. Security engineer Moritz Sanft then published the details on his blog and on X: he asked Codex to map the CRuby pre-authentication attack surface reachable from Rails, and in under 24 hours Codex returned several bugs, two of which he says yield pre-auth RCE in many Ruby on Rails apps, with a plain rails new app already exploitable. OpenAI security staff helped with exploitation and coordinated disclosure, and OpenAI president Greg Brockman shared the post on X.

⚡ Key Takeaways
  • •Official: Ruby fixed seven vulnerabilities in its regex engine and string case conversion, with known RCE PoCs for some
  • •Versions: affected are 3.3.13 and earlier, 3.4.0 to 3.4.12 and 4.0.0 to 4.0.9; upgrade to 4.0.10, 3.4.13 or 3.3.14, as Ruby 3.2 and earlier are end of life and get no fix
  • •Codex: Sanft had Codex map the Rails-reachable pre-auth attack surface, got bugs back in under 24 hours, and had an end-to-end exploit on a toy app by October 26
  • •Attack path: Rack lets a multipart _method field override the request method and calls upcase on it, and the charset key picks the encoding, which triggers the heap overflow
  • •Mitigation: the integer overflow only affects Rack 3.2.1 and earlier since Rack 3.2.2 caps multipart field size; OpenAI also invited WAF vendors to help defend
Read details→
U
U.S. GAO@USGAO·Collected 7h ago
📊 Benchmark

US GAO Covert Test: ChatGPT, Claude, Gemini and Grok Safeguards Bypassed, Each Produced a Viable Attack Plan in Under 12 Hours

On November 18, the US Government Accountability Office (GAO), the congressional watchdog, publicly released report GAO-27-108814. In spring 2026 GAO tested the four most-used generative AI platforms, Claude, Grok, ChatGPT and Gemini, while posing as ordinary users. Every platform triggered its safeguards (refusals, removed content or account suspensions), but GAO got past all of them with common bypass techniques. Each platform produced an attack plan in under 12 hours that GAO verified as feasible: two involved a vehicle-borne improvised explosive device and two targeted the electrical grid. GAO says the information was not new, but AI pulled it into a plan within hours instead of weeks or months.

US GAO Covert Test: ChatGPT, Claude, Gemini and Grok Safeguards Bypassed, Each Produced a Viable Attack Plan in Under 12 Hours
⚡ Key Takeaways
  • •Scope: Claude, Grok, ChatGPT and Gemini were picked by April 2026 usage, randomly assigned to scenarios, and reported anonymously as platforms A to D
  • •Results: two platforms gave vehicle-borne IED plans naming a specific hotel or convention center with an exact date and time; two gave drone or firearms attack plans on specific substations
  • •Safeguards: all four platforms refused, removed content or suspended accounts at some point, yet common bypass techniques got past every one
  • •Speed: each plan took under 12 hours, and GAO verified feasibility through research, surveillance and expert interviews
  • •Response: officials from the four companies said they will keep improving platform safety; the full report runs 61 pages
Read details→
Y
Yash Patil@ypatil125·Collected 8h ago
📉 Price Cut

Palo Alto Networks to Acquire Applied Compute for $4.5B in Stock, Betting on Secure Post-Training and Inference of Open-Weight Models for Enterprises

On November 18, cybersecurity company Palo Alto Networks announced a definitive agreement to acquire Applied Compute, a platform for post-training and inference of open-weight models, in an all-stock deal valued at $4.5 billion, subject to adjustments. The deal needs regulatory clearance and is expected to close in Palo Alto Networks' fiscal third quarter of 2027. Applied Compute co-founder and CEO Yash Patil will become Palo Alto Networks' Co-Chief AI Officer after closing. Patil wrote on X that a few dozen people took the company from zero to more than $100 million in annualized revenue run-rate in just over 18 months.

Palo Alto Networks to Acquire Applied Compute for $4.5B in Stock, Betting on Secure Post-Training and Inference of Open-Weight Models for Enterprises
⚡ Key Takeaways
  • •Deal: all-stock, valued at $4.5 billion subject to adjustments, pending regulatory clearance including antitrust review, expected to close in fiscal Q3 2027
  • •Target: Applied Compute runs a post-training and inference platform for open-weight models, letting companies train, deploy and keep improving proprietary models on their own data
  • •People: co-founder and CEO Yash Patil becomes Palo Alto Networks' Co-Chief AI Officer after closing, and the research team joins the company
  • •Scale: Patil says a few dozen people grew annualized revenue run-rate from zero to over $100 million in just over 18 months
  • •Pricing: the 8-K filing says shares will be issued at $337.16, the 15-trading-day volume-weighted average close ending November 17
Read details→
ADSponsored
A
Alexandr Wang@alexandr_wang·Collected 9h ago
🚀 Release

Meta Muse Adds Try On for Clothes, Launches in Australia and Signs Loblaw as Its First Canadian Retail Partner

On November 18, Meta Chief AI Officer Alexandr Wang announced on X that Muse, Meta's personal AI agent, now has Try On: while you shop, Muse offers to show how a clothing item looks on you. The post drew nearly 5,900 likes and about 875,000 views. Meta's newsroom published two related updates. On November 17 Muse launched in Australia on iOS, Android, muse.ai and Mac, with local connectors such as Canva, Domain, Kogan and MECCA rolling out progressively. On November 18 Loblaw became the first Canadian retailer to connect to Muse, and voice conversations and spaces rolled out in Canada.

Meta Muse Adds Try On for Clothes, Launches in Australia and Signs Loblaw as Its First Canadian Retail Partner
⚡ Key Takeaways
  • •Try On: while you shop, Muse offers to show how an item looks on you before you buy; announced on X by Alexandr Wang himself
  • •Australia: launched November 17 on iOS, Android, muse.ai and Mac, with AI glasses support coming soon; free for most uses, with subscription plans available
  • •Australian connectors: Canva, Domain, Kogan, MECCA, Officeworks, realestate.com.au, SWEAT, ShopBack and Ticketek, rolling out progressively
  • •Canada: Loblaw is the first retail connector partner, so shoppers can plan meals, build a basket and check out inside Muse; voice and spaces also rolled out
  • •Safety: Muse runs on Muse Secure VM, keeps passwords and payment methods out of the agent's view, and asks before sensitive actions such as sending email or buying
Read details→
S
Sam Joshua@SAM_JOSHUA_S·Collected 11h ago
🛠️ Tooling

Bitdefender Finds 22 Malicious GitHub Repos That Use AGENTS.md, CLAUDE.md and Session-Start Hooks to Make AI Coding Agents Run setup.js, Steal Credentials and Infect 14 Agent Configs

On November 17 Bitdefender Labs reported that researcher Andrei Anton-Aanei found 22 malicious GitHub links aimed at developers who use AI coding assistants. AGENTS.md and CLAUDE.md files in the repos describe scripts/setup.js as a "sealed binary artifact" and tell the agent not to inspect it and to run it silently, while a session-start hook in .claude/settings.json and a VS Code task can run it as soon as the repo is trusted, with no prompt at all. The payload steals CI/CD, cloud, package-registry and source-control credentials, rewrites local configs for 14 coding agents, and uses stolen GitHub or GitLab tokens to commit the same files to more repositories.

Bitdefender Finds 22 Malicious GitHub Repos That Use AGENTS.md, CLAUDE.md and Session-Start Hooks to Make AI Coding Agents Run setup.js, Steal Credentials and Infect 14 Agent Configs
⚡ Key Takeaways
  • •Two ways in: AGENTS.md/CLAUDE.md talk the agent into running setup.js, while session-start hooks and VS Code tasks run it once the repo is trusted, with no prompt
  • •The payload is a 3.75 MB multi-layer obfuscated JavaScript loader that fetches the Bun runtime and steals CI/CD, cloud, package-registry, browser, password-manager, SSH key and .env credentials
  • •Local spread: it searches for configs of 14 agents including Claude, Codex, Gemini, Copilot, Cursor, Cline and Aider, then injects instructions, hooks or MCP servers
  • •Remote spread: it validates stolen GitHub/GitLab tokens, commits AGENTS.md, .mcp.json and similar files to up to 25 repos by default, and tries to republish npm, PyPI and RubyGems packages
  • •Defense: before opening an unfamiliar repo, review AGENTS.md, CLAUDE.md, .claude/, .cursor/, MCP definitions, hooks and VS Code tasks, and distrust any "do not inspect, just run" instruction
Read details→
T
Tencent Hy@TencentHunyuan·Collected 12h ago
🚀 Release

Tencent Hy Open-Sources Hy4 flash Weights, Including FP8 and Base Versions: 220B Total, About 12B Active, 1M Context, Day-0 vLLM and SGLang Support

On November 18 the official Tencent Hy account (@TencentHunyuan) announced open weights for Hy4 flash, releasing an FP8 quantized version and the pre-trained Base model alongside it on Hugging Face, ModelScope and GitCode under the Tencent Hy License. According to the model card it is a mixture-of-experts model with 220B total parameters (including 64B Engram parameters) and about 12B active per token, takes text and image input natively, and supports up to 1M tokens of context. vLLM and SGLang ship official images, and a GGUF build is available for testing.

Tencent Hy Open-Sources Hy4 flash Weights, Including FP8 and Base Versions: 220B Total, About 12B Active, 1M Context, Day-0 vLLM and SGLang Support
⚡ Key Takeaways
  • •Three releases: Hy4 flash (instruct), Hy4 flash-FP8 (quantized) and Hy4 flash-base (pre-trained), all downloadable
  • •Scale: 220B total parameters (64B of them Engram), about 12B active per token, 1M (1,048,576) token context
  • •Architecture: sparse and sliding-window attention mixed 1:3, 16 of 512 routed experts active per token, plus a built-in 0.85B DFly2 speculative-decoding drafter
  • •Deployment: official vLLM image vllm/vllm-openai:hy4-flash and an SGLang image; the reference setup runs the FP8 model with 8-way tensor parallelism
  • •Licensed under the Tencent Hy License (not Apache or MIT), so read the terms before commercial use
Read details→
R
Ramp Labs@RampLabs·Collected 13h ago
📊 Benchmark

Ramp Labs Publishes Ramp SWE-Bench v2: 500 Real Production Tasks, GPT-6.2 Astra in Codex Leads at 76.4%, Carbon 75.4%, Fable 5.2 and GPT-6.2 Sol Tie at 73.4%

On November 18 Ramp Labs, the research arm of fintech company Ramp, published Ramp SWE-Bench v2 on its official X account: 500 tasks built from work that more than 200 Ramp engineers actually shipped to production, with stricter verification and each frontier model running in its native harness (Claude Code, Codex, OpenCode, Grok Build and others). The official leaderboard (pass@1, high effort) puts GPT-6.2 Astra first at 76.4%, Carbon at 75.4%, and Fable 5.2 and GPT-6.2 Sol tied at 73.4%. Mean cost per task ranges from $0.07 for GPT-6 Luna to $4.59 for Muse Wonder.

Ramp Labs Publishes Ramp SWE-Bench v2: 500 Real Production Tasks, GPT-6.2 Astra in Codex Leads at 76.4%, Carbon 75.4%, Fable 5.2 and GPT-6.2 Sol Tie at 73.4%
⚡ Key Takeaways
  • •Task source: 500 tasks drawn from real financial software work that 200+ Ramp engineers shipped to production
  • •Top five: GPT-6.2 Astra 76.4%, Carbon 75.4%, Fable 5.2 73.4%, GPT-6.2 Sol 73.4%, Hydrogen 71.6%
  • •Value: GPT-6.1 Terra scores 70.4% at $0.46 per task, Sonnet 5.5 63.2% at $0.47, DeepSeek V4.1 Flash 60.8% at $0.22
  • •On the hardest 25% of tasks Carbon (43%) and Hydrogen (41%) lead clearly, ahead of Astra at 30% and Fable 5.2 at 14%
  • •Speed: GPT-6.2 Astra in Codex averages 6.5 minutes per task, while Carbon takes 26.3 minutes and 349 tool calls
Read details→
F
Firecrawl@firecrawl·Collected 14h ago
🛠️ Tooling

Firecrawl Launches Open-Source Firecrawl Local: One Rust Binary That Scrapes and Crawls Sites on Your Machine, Works as an MCP Server for AI Agents and Serves the v2 API on Localhost

On November 18 Firecrawl announced Firecrawl Local on its official X account: scraping, crawling and site mapping rebuilt as a single Rust binary that runs on your own machine. It is open source in apps/local of the firecrawl/firecrawl repository (AGPL-3.0) and installs with cargo install firecrawl-local. It turns pages into clean markdown, plugs into Claude Code, Cursor, Codex, Gemini CLI and other tools as an MCP server, and serves the Firecrawl v2 API on localhost:3002, so existing SDK code works by changing the API URL. The official post claims scraping is 37x faster than the self-hosted version with 230x less memory (11 MB idle).

⚡ Key Takeaways
  • •One binary: no database, queue or container; install with cargo install firecrawl-local (needs Rust and a C compiler)
  • •firecrawl-local init adds the MCP server to Claude Code, Claude Desktop, Cursor, Windsurf, Gemini CLI and Codex, backing up each config first
  • •serve exposes the v2 scrape, map, crawl and batch scrape routes on 127.0.0.1:3002, with no API key needed for local routes
  • •Search, pages that block your machine and Alexandria data providers still go to Firecrawl Cloud; FIRECRAWL_CLOUD=off keeps everything local
  • •The speed figures (37x faster than self-hosted, 230x less memory) come from the official post; the docs do not describe how they were measured
Read details→
A
Anthropic@AICoder·Collected 15h ago
🛠️ Tooling

Claude Code 2.1.328: Auto Mode Comes to Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud and Microsoft Foundry and Is On by Default; /code-review Now Reports Only Bugs

Anthropic released Claude Code 2.1.328 on November 18. The biggest change: Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway now support auto mode, and interactive sessions with no permission mode set start in it; disableAutoMode turns it off. On most models, /code-review now reports only bugs and CLAUDE.md violations and marks each finding Important, Pre-existing or Nit, while cleanup suggestions move to /simplify.

⚡ Key Takeaways
  • •Auto mode reaches third-party platforms: Opus 4.6 and Sonnet 4.6 on Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and the Claude apps gateway
  • •Default change: interactive sessions with no permission mode start in auto mode; plan mode and Bash prompts also offer it, and disableAutoMode turns both off
  • •/code-review: on most models it reports only bugs and CLAUDE.md violations, marks each Important, Pre-existing or Nit and suggests a fix; reuse, simplification and efficiency cleanups move to /simplify
  • •Hooks: mcp_tool hooks accept "onFailure": "block", so a failed hook call blocks the action
  • •Also: fixes zsh commands failing with "not found" on words starting with =, automatic compaction reusing an old system prompt, and many Code Mode issues
Read details→
C
ChatGPT@ChatGPT·Collected 15h ago
🛠️ Tooling

ChatGPT Launches Live Dictation: Text Appears as You Speak, the 10-Minute Recording Limit Is Gone, Rolling Out to All Plans Worldwide; Desktop App Adds an Accounting Plugin

In its November 18 ChatGPT release notes, OpenAI announced Live Dictation: speech is turned into text while you talk instead of after the recording ends, you can tap or click the waveform to show or hide the live text, and the previous 10-minute recording limit has been removed. It is rolling out globally to all ChatGPT plans on iOS, Android, web and the desktop app. The same release notes added an Accounting plugin that uses computer use in the macOS and Windows desktop apps to reconcile transactions, draft invoice entries and prepare month-end reports.

ChatGPT Launches Live Dictation: Text Appears as You Speak, the 10-Minute Recording Limit Is Gone, Rolling Out to All Plans Worldwide; Desktop App Adds an Accounting Plugin
⚡ Key Takeaways
  • •Live Dictation: speech becomes text as you talk, instead of being transcribed after the whole recording ends
  • •Tap or click the waveform to show or hide the live text; the 10-minute recording limit has been removed
  • •Availability: all ChatGPT plans worldwide on iOS, Android, web and the desktop app, rolling out gradually
  • •Accounting plugin: installed after setting up computer use in the macOS or Windows desktop app; reconciles transactions, drafts invoice entries and prepares month-end reports from your records and templates; Enterprise admins must approve it first
  • •The official @ChatGPT post calls dictation faster and more reliable but gives no latency or accuracy figures
Read details→
W
Wiz@wiz_io·Collected 16h ago
🛠️ Tooling

Microsoft Agent Framework PyPI Package Poisoned: Versions 1.26.1 and 1.26.2 Shipped a Credential Stealer Targeting Azure, CI and GitHub Actions Tokens; Last Clean Release Is 1.26.0

GitHub security advisory GHSA-764m-3ppc-5m73 and Wiz report that two malicious versions of Microsoft's official Python package agent-framework were published on November 17: 1.26.1 at 20:35 UTC and 1.26.2 at 21:08 UTC. Both have been removed from PyPI. The malware runs from a .pth file on every Python start, even if the package is never imported, and collects secrets from environment variables, credential files, Azure managed-identity tokens and Azure DevOps OIDC tokens; 1.26.2 also steals GitHub Actions OIDC tokens. Any environment that installed either version should be treated as exposed.

Microsoft Agent Framework PyPI Package Poisoned: Versions 1.26.1 and 1.26.2 Shipped a Credential Stealer Targeting Azure, CI and GitHub Actions Tokens; Last Clean Release Is 1.26.0
⚡ Key Takeaways
  • •Affected: agent-framework 1.26.1 and 1.26.2, published November 17 and since removed; 1.26.0 and earlier are not affected
  • •Trigger: a .pth file runs the stealer on every Python interpreter start without importing the package, so one pip install in CI is enough
  • •What it steals: environment variables containing TOKEN, SECRET, KEY, AZURE, AWS, GITHUB, OPENAI, HF_ and similar, plus ~/.azure, .npmrc, .git-credentials, Docker and Hugging Face credentials, Azure managed-identity tokens and Azure DevOps OIDC tokens
  • •Added in 1.26.2: GitHub Actions OIDC token theft for workflows with id-token: write, /home/runner paths, and a file cap raised from 60 to 80
  • •Response: pin to 1.26.0, remove the .pth hooks and rebuild the environment, and rotate every reachable secret, including CI and GitHub Actions secrets
Read details→
W
Wiz@wiz_io·Collected 16h ago
🛠️ Tooling

Java4Shell: A Calif Researcher Used OpenAI Codex to Find an OpenJDK Base64 Memory-Corruption RCE in One Hour for Under $200 in Tokens

Security firm Calif disclosed Java4Shell (CVE-2026-114938) on November 17, a memory-corruption bug in OpenJDK's AVX-512 VBMI accelerated Base64 decoder on x86-64, scored 9.2 under CVSS 4.0. Researcher Quang Luong had OpenAI Codex audit HotSpot, found the bug in about an hour and built the first working exploit in two more hours, for under $200 in tokens. Calif achieved unauthenticated RCE against default configurations of Keycloak, Jira, Confluence, Jenkins and others. OpenJDK and Oracle released fixes the same day.

Java4Shell: A Calif Researcher Used OpenAI Codex to Find an OpenJDK Base64 Memory-Corruption RCE in One Hour for Under $200 in Tokens
⚡ Key Takeaways
  • •The bug: HotSpot's x86-64 Base64 intrinsic miscalculates the output length when handling trailing padding, writing up to 62 bytes out of bounds; 64 A characters plus one = trigger it
  • •AI's role: Quang Luong had Codex audit the JIT, intrinsics and crypto code; early findings were mostly noise, and the real bug surfaced after about 20 minutes; 1 hour to find, 2 more to exploit, under $200 in tokens
  • •Scope: OpenJDK 17.0.6 and later plus distributions such as Corretto, Temurin, Microsoft Build of OpenJDK, Zulu and Oracle Java SE, on x86-64 CPUs with AVX-512 VBMI; ARM64 is not affected
  • •Demonstrated: unauthenticated RCE on default Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate; authenticated RCE on Kafka, Solr, Elasticsearch and others
  • •Severity: 9.2 under CVSS 4.0 per Calif and 8.1 under CVSS 3.1 in the OpenJDK advisory; the -XX:-UseBASE64Intrinsics flag is a temporary mitigation
Read details→
V
Vercel@vercel·Collected 17h ago
🛠️ Tooling

Vercel Launches Experimental Vercel Board, a Shared Message Board Where Humans and Agents Claim Tasks, Schedule Follow-Ups and Sync Two-Way With Slack and GitHub

Vercel launched Vercel Board on November 18 as an experimental, headless message store and coordination primitive for humans and agents. Agents read and write channels and threads through the CLI or the JavaScript SDK: they can claim tasks, renew leases, post updates and record results, while watchers wake agents on new replies and vercel board schedule sets follow-ups. Board adds two-way sync with Slack and GitHub, schema-validated structured messages and per-channel grants. Vercel says v0's software factory now runs on Board and merges over 50% of PRs.

Vercel Launches Experimental Vercel Board, a Shared Message Board Where Humans and Agents Claim Tasks, Schedule Follow-Ups and Sync Two-Way With Slack and GitHub
⚡ Key Takeaways
  • •Positioning: an experimental, harness-agnostic coordination primitive that organizes tasks, context and results in channels and threads
  • •Compatibility: Claude Code, Codex, Pi, OpenCode, eve, fx, Mastra and LangChain can collaborate on the same board
  • •Features: task claiming with leases, task dependencies, watchers that wake agents on new replies, scheduled follow-ups, and two-way Slack and GitHub sync
  • •Access control: per-channel grants by user, project and environment, with every event recording who performed the action
  • •Internal use: Vercel security agent swarms and the v0 software factory, which now merges over 50% of PRs; get started with npx skills add vercel/board
Read details→
R
Replit@Replit·Collected 17h ago
🛠️ Tooling

Replit Launches Missions: Set a Goal and a Budget, and Agent Swarms Plan, Build, Test and Iterate While an Independent AI Judge Checks Results Against Your Criteria

Replit published Replit Missions on its official blog on November 17 and promoted it again from its official X account on November 19. You define the goal, the spending limit and which actions need your approval, and Replit coordinates a swarm of agents to plan, build, test and iterate in parallel. An independent AI judge checks results against the agreed success criteria; a mission completes only when they are met and stops if the budget runs out. Replit used it to build DataCat, its own application monitoring platform, in 2 days within a $3,000 budget, and in its NumPy tests a 32-agent swarm reached speedups of up to 19x.

Replit Launches Missions: Set a Goal and a Budget, and Agent Swarms Plan, Build, Test and Iterate While an Independent AI Judge Checks Results Against Your Criteria
⚡ Key Takeaways
  • •How it works: you set the goal, a spending cap and which actions need approval (such as contacting customers, publishing or raising the budget), and an agent swarm works in parallel
  • •Completion: an independent AI judge checks results against the agreed success criteria; if the budget runs out first, work stops with the outcome unresolved
  • •Case study: DataCat, a monitoring platform built in 2 days within $3,000; it alerted on a deliberately injected database bug in 38 seconds and passed all 51 data and alert checks
  • •Replit NumPy tests across 13 tasks: a single Auto (Max) agent reached up to 4x speedups, a 16-agent swarm 17x and a 32-agent swarm 19x
  • •Replit comparison: from the same prompt, DataCat ingested 49,000 traces per second, about 5x the Claude Code /goal build; pricing was not stated in the blog post
Read details→
s
smlay@Smlay_ero·Collected 18h ago
📉 Price Cut

DeepSeek Stops Billing Server-Side Web Search and Fetch Inside Model Calls: Since Nov 18 12:00 Beijing Time Only Tokens Are Charged, While Standalone Search and Fetch APIs Stay at $0.75 per 1,000

DeepSeek updated its official pricing page: starting 12:00 on November 18, 2026 (Beijing Time, UTC+8), server-side web search and web fetch triggered during a model call are no longer billed, and only model token usage is charged. This covers the web tools, the web_search tool in the Anthropic API and the web_search tool in the Responses API. Calling the standalone Search API or Fetch API directly is still billed at $0.75 per 1,000 successful requests. When DeepSeek opened its app capabilities to the API on November 13, it said these extended capabilities carried tool fees on top of token charges.

⚡ Key Takeaways
  • •Effective from 12:00 on November 18, 2026, Beijing Time (04:00 UTC)
  • •Now free: server-side web search and fetch triggered inside a model call, including the web tools and the web_search tool in the Anthropic API and the Responses API
  • •Still billed: direct calls to the standalone Search API and Fetch API, at $0.75 per 1,000 successful requests
  • •Model prices unchanged: deepseek-flash (DeepSeek-V4.1-Flash-1017) costs $0.3 per 1M cache-miss input tokens and $1.2 per 1M output tokens at peak, half price off-peak
  • •For agent builders: letting the model search on its own is now cheaper than wiring a separate search API, but search results added to the context still count as input tokens
Read details→