Highlights
What happened at the companies we cover, by day, newest first. Every item links to its sources.
Showing 1–20 of 746 highlights
Saturday, Nov 14 18 stories
Users show Replit Agent running Google Ads on its own as approval questions surface
DiscussionX
Two days after Replit gave its Agent the ability to run Google Ads and buy supplies through Stripe's Link and Visa, early users began posting results on X. One builder said asking Replit to grow an app led Agent to set up a Google Ads account, start campaigns and generate tens of thousands of paid impressions within 48 hours, and CEO Amjad Masad amplified it as autonomous businesses arriving. Replit's blog says Agent stops at checkout for owner approval, with approval checks enforced in code, and Replit Support repeated in replies that Agent prepares checkout for the user to review and approve. Reaction was largely enthusiastic, while one reply asked that each charge record show who approved it and for how much.
Sources
2 more sources
Cursor breach report spreads; developers told to check extension versions and use Privacy Mode
DiscussionNewsXHNBluesky
Cursor's Nov 13 incident update spread widely on Saturday across security press, X, Bluesky and a small HN thread. The report says a key Cursor issued to Fireworks was exposed, letting an attacker copy mostly 2025 agent traces from non-Privacy-Mode sessions, under 0.05% of users, some holding unredacted credentials. Those led to internal tools and a credential used to plant tampered Claude Code builds 2.1.312 and 2.1.313 in Cursor's marketplace cache for about 11 hours, reaching at most 5,000 users. BleepingComputer adds the malware installed persistent SSH backdoors; Cursor says it moved all inference off Fireworks and rotated credentials. Reaction was largely negative, with posts urging version checks and Privacy Mode and calling stored agent logs an attack surface.
Sources
2 more sources
Instinct users keep weighing it against Muse and Dot, with trust and suspension gripes
DiscussionX
Instinct's personal agent kept coming up on X on Saturday as people compared it with rivals such as Muse, Dot and Duet. The most-engaged post, from a Korean user, praised Instinct as a strong agent that works over iMessage without installing anything and said they run it alongside Muse and another agent. Other users said they were confused about how the competing agents differ, while one argued Muse, Dot and Instinct already cover what a newer entrant offers. On the skeptical side, a user called handing an AI agent bank-account access horrifying, and another said their Instinct account was suspended with no explanation and no reply from support.
Sources
2 more sources
LangGraph 1.2.22 ships a batch of runtime performance fixes and a subgraph streaming fix
LaunchOther
LangChain published LangGraph 1.2.22 late on Nov 13 PT alongside new checkpoint, checkpoint-postgres and prebuilt package releases. The release notes list several runtime performance changes, including skipping message conversion in add_messages for plain messages, matching update keys against a set on wide states, and filtering pending writes before reading task statuses. It also fixes streaming when a child graph runs on its own thread while the caller streams subgraphs. The update is incremental and drew little public discussion, landing as social chatter about LangChain stayed focused on its Series C.
Sources
Developers keep calling Devin's plans with free SWE-2 the best-value coding subscription
PraiseX
Through Saturday, many developers on X in English, Chinese, Japanese and Korean urged others to sign up for Devin while its in-house SWE-2 model is free on paid plans, some saying they would make it their main coding agent or drop another subscription. One widely shared post ranked Devin's and Claude's $20 plans as the best value on the market, and another highlighted that Pro users now also get free SWE-2 in concurrent Devin cloud sessions. Not everyone agreed: a Chinese-language roundup said group feedback was split between praise for speed and cloud sandboxes and complaints that SWE-2 trails top models and that third-party model quotas are small. One user also reported SWE-2 Flash could not read images in the CLI.
Sources
2 more sources
Devin's Nov 13 release adds an artifacts rail, review analytics and GitLab merge trains
LaunchOther
Cognition's Nov 13 Devin release notes rework the session layout: pull requests, diffs, previews, plans and other outputs now sit in a compact rail beside the chat instead of opening tabs on their own, and the Agent/Ask switch moves into the composer with per-side modes. Devin Review gains an analytics view showing whether findings were fixed, rejected or ignored, broken out by each custom reviewer and each version of its instructions. Review can also add GitLab merge requests to a project's merge train, and the session file editor now saves only on Cmd/Ctrl+S and tells Devin which files you saved. Smaller changes include a German UI, clearer errors when a code host is slow, and the option to unlink Devin from a security finding.
Sources
Hub takedown scare on X turns into a debate about mirroring open weights off Hugging Face
DiscussionX
A viral X post on Saturday claimed Hugging Face had pulled uncensored model repos without warning and urged people to download everything. Its author later corrected that, saying a Hugging Face leader explained the repo was removed for abusing the gating system to take users' personal data, not for being uncensored. The correction still argued that relying on one hub for nearly all open weights is risky with the Nvidia acquisition pending and Huang gone, and recommended local copies, mirrors or torrents alongside HF. Others pointed out HF had already removed a model marketed for offensive cyber use while press tallies of thousands of abliterated models frame them as a risk. Reaction mixed praise for HF's team with unease about concentration of open-weight distribution.
Sources
Engineers debate per-model inference engines after Baseten's from-scratch DeepSeek engine
DiscussionX
Discussion of Baseten's open-source Rust/CUDA engine for DeepSeek V4.1-Flash kept building on X through Saturday. A follow-up thread from the team said the build took about four days, six prompts and roughly $500 in tokens, beat SGLang by about 15%, and that the optimizations were then ported back and upstreamed to SGLang. Reactions split: some praised the work and argued model-specific engines are worth the fragmentation, others framed it as proof that agent-written infrastructure is now cheap. Skeptics countered that the hardware economics, around $140K for 1K tokens per second, keep this out of reach for local or indie hosting.
Sources
2 more sources
Vercel Daemons can now export OpenTelemetry traces to Vercel Tracing
LaunchBlueskyOther
Vercel's Nov 13 changelog says background Daemons can now ship OpenTelemetry traces into Vercel Tracing, next to the rest of a project's traces. Each daemon gets a Unix socket for OTLP/HTTP, exposed through an environment variable, and @vercel/otel 2.3.0 or later picks it up automatically for Node.js workers. Spans can be searched by service name from the Vercel CLI with JSON output, which Vercel pitches as something agents can query directly. The update was picked up in Japanese-language developer posts on Bluesky on Nov 14, with little debate.
Sources
Report on Lovable apps leaking Supabase data draws a 700-point HN thread
DiscussionHNX
Hamburg security firm Crissy Field's research on Lovable-built apps drew a large Hacker News thread on Saturday, passing 700 points and nearly 300 comments; the findings first surfaced in early November via Handelsblatt and Cybernews. The firm checked 159,041 Lovable projects found in Common Crawl, saw 24,664 Supabase-backed sites return records without any login, and found email-bearing tables on 7,403 of them, cautioning that this is not a confirmed leak rate. Exposed records included passport details, more than 200,000 forum chat messages, ticket and payment data and contracts; the researchers point to missing database access rules, not a breach of Lovable itself. HN commenters split blame among vibe coding without security know-how, Supabase's defaults and a lack of liability, while X posts repeated the numbers and urged builders to check row-level security.
Sources
1 more source
Baseten launches a Voice Agent Stack and opens a Realtime Voice private preview
LaunchNewsOther
Baseten introduced a Voice Agent Stack for running low-latency production voice agents, placing the orchestrator and speech and language models together on a tight in-cluster network. Alongside it, the company opened a private preview of Baseten Realtime Voice, a reference agent that supports open-source speech models and the OpenAI Realtime protocol at roughly one cent per audio minute. A community writeup notes teams can bring and swap their own speech-to-text, LLM and text-to-speech models, with components scaling independently. Baseten's signup page says it is choosing a small set of design partners, and public reaction so far has been light.
Sources
Builders pick apart Poppy 1.0 a day after Sierra's agent protocol hit every Shopify store
DiscussionXBlueskyNews
The day after Sierra shipped version 1.0 of its Personal Agent Protocol (Poppy), developers on X and Bluesky spent Saturday unpacking how it works. Most posts focused on the mechanics Sierra described: each Shopify store now auto-publishes a Poppy file, Meta's Muse looks for it first, and customers scope agent access via OAuth on top of MCP, OpenAPI and UCP. Sierra's release also lists 60 new design partners, taking the total past 100, with Mercury, Synchrony and Turkish Airlines going live. Reaction was mostly positive about a sanctioned entry point replacing site scraping, though some replies asked how people will see and cap what an agent shares or buys, and a community Sylius plugin implementing the spec already appeared on Bluesky.
Sources
2 more sources
Praxis-1 open release draws scrutiny over self-run benchmarks and robot-data needs
DiscussionNewsXBluesky
A Nov 14 analysis by The Decoder walked through Runway's newly opened Praxis-1 robot model, released with weights and code under Apache 2.0. It notes Runway reports 46.2% on Nvidia's RoboLab-120, top among world action models, but that Runway ran all tests itself and results are not independently verified. The piece adds that adapting to a Franka arm still took about 277 hours of DROID recordings, that the model can confuse similar objects and cannot reliably plan long multi-step tasks, and that the standard build needs an 80 GB GPU. On X and Bluesky, many posts recapped the release, with some explicitly flagging the self-reported benchmarks and the data requirements.
Sources
Producers weigh Suno Bridge's requirement to upload Ableton sessions to Suno's servers
DiscussionXNews
A day after Suno released the Bridge beta for Ableton Live 12 on Mac, much of the X conversation shifted from the feature itself to how it works. Suno's own post confirms that syncing a session means uploading its files to Suno so the models can read tempo, key and harmony, with generated parts landing on the Live timeline. A trade writeup adds that only the chosen project is sent, uploads are filtered for copyrighted material, and generation spends the same credits as Suno Studio. Creator posts mostly liked keeping their own plugins and routing but flagged the cloud upload and Premier-only access as the catch, so reaction was mixed rather than hostile.
Sources
1 more source
Cursor traces its breach and tampered Claude Code extension to a key exposed at Fireworks
NewsNewsXBlueskyHN
Cursor's incident report says an attacker obtained an access key Cursor had issued to Fireworks, which Fireworks exposed, and used it to copy a shared bucket of mostly 2025 coding-agent traces from non-Privacy-Mode sessions. Some traces held unredacted credentials, which led to Cursor SaaS tools and a credential used on Oct 29 to plant malicious Claude Code extension builds that up to 5,000 users may have received over about 11 hours. Cursor says it has moved all inference traffic off Fireworks, cut its access to shared datasets and rotated credentials. BleepingComputer covered the disclosure on Nov 14, placing it within the wider October wave of attacks on AI dev platforms, and the fallout was widely shared on X, Bluesky and HN as another downstream customer hit by the Fireworks breach.
Sources
Perplexity Agent API adds Gemini 4 Hydrogen and unlists three older Gemini Flash models
LaunchOther
Perplexity's API changelog now lists support for google/gemini-4-hydrogen in its Agent API. The model is priced at $1.50 per million input tokens, $0.075 per million cached input tokens and $7.50 per million output tokens, with reasoning tokens billed as output. In the same entry, gemini-3.6-flash, 3.7-flash and 3.8-flash were dropped from the model docs, pricing calculator and examples, though the changelog does not say calls to them will fail. A docs-tracking site caught the change on Nov 14 and noted Hydrogen undercuts the Gemini 4 Carbon pricing added days earlier.
Sources
Profound's AI Marketer agent can now be emailed and keeps visible memories
LaunchOther
Profound's Nov 13 changelog gives its AI Marketer assistant a company email address: customers can forward messages to it, CC it quietly on threads, and get answers back by email, with the exchange also logged in the app. The assistant and Profound Agents can also send email to anyone on the customer's company domain without connecting a mailbox, for example a recurring competitive brief. The same release makes the assistant save preferences to a Personal Context store during chat, shows a chip for each saved memory, and lets it draw on connected tools such as Notion, Google Drive, Slack or Jira. Each memory links back to its source. There was little outside discussion of the update on Saturday.
Sources
Huang's exit puts Nvidia's pending Hugging Face deal and dealmaking pace in question
NewsNewsXBluesky
Coverage on Saturday examined what Jensen Huang's resignation means for Nvidia's pending purchase of Hugging Face, agreed in September and expected to close in the first half of 2027 subject to regulatory approval. With Colette Kress as interim CEO and a board search underway, reports suggest Nvidia's aggressive startup dealmaking is likely to slow, and The Information said Huang personally drove deals like the HF acquisition, often favoring speed over diligence. HF CEO Clément Delangue has said the company will keep fighting for open-source AI regardless of Nvidia's leadership, and Nvidia has said HF will stay open without requiring Nvidia hardware. On X and Bluesky, commentators questioned whether the deal still makes sense and what happens to the main open-weights hub; reported deal values differ between $12.3B and $12.9B.
Sources
1 more source
Friday, Nov 13 42 stories
Three major outages hit Kimi K3.1 Pro on Together AI in one day
ComplaintXBluesky
Together AI's status feed reported three separate major outages for its Kimi K3.1 Pro endpoint on Nov 13. The incidents were marked resolved after roughly 36 minutes, 3 minutes and 21 minutes respectively, with no resolution time given while the third was open. Monitoring accounts relayed each start and resolution on X and Bluesky, though there was little direct user commentary. Three failures in one day on a single popular open-weight model is a reliability signal for teams depending on that endpoint.
Sources
2 more sources
Gamma details how splitting its agent into jobs let small models cut deck cost to cents
OtherXOther
Gamma co-founder Jon Noronha published the first Gamma Research post in November, explaining how the Gamma 5 agent moved from one frontier model, which cost $1.30 to $7.40 per deck in early prototypes, to many small models handling separate jobs such as planning, template choice, writing, repair and review. Slides are rendered in a browser and checked by hundreds of programmatic rules, with a model called back only when a check fails, and each job is A/B tested on live traffic, with more than twenty model experiments in October. Gamma found no single best small model, uses Haiku for edits and GPT-6 Luna for slide repair, and says a small repair model it trained fixed 76% of failing slides on the first try versus 57% for a general model. The post spread on X with positive reactions focused on the cost drop.
Sources