RuntimeWire

Featured

• Amazon talks to buy Decart for $7B as chip software draws a premium

  The reported price would top Decart's $4 billion valuation in May, when it raised $300 million led by Radical Ventures.

• Yandex says drone attack halted its Vladimir data center operations

  A Bluesky user claimed bot traffic from Yandex's network had fallen near zero after the strike. Cloudflare Radar's accessible chart and a linked screenshot do not establish the drop's interval, baseline or magnitude.

• Startup Spotlight: Parrot pivoted into collision repair and signed two $30,000 shop contracts

  Parrot founders Phillip Baek and Liam McBride pivoted from biomedical workflow automation three weeks before YC Demo Day, then secured two $30,000 annual shop contracts.

• Startup Spotlight: Mesmer sends engineering fixes into customer codebases

  Joao de Paula and his co-founders are moving from engineering reports to pull requests, betting that teams need help fixing bottlenecks as AI increases the volume of code.

• Startup Spotlight: Tergle builds custom AI audit agents for large-company workflows

  Adam Bowen and William Tarr say their agents automate repetitive audit procedures while keeping people in the review loop. Tergle says it works with multiple billion-dollar companies, but has not named them; its disclosure-review speed claim is founder-reported.

• HeyGen ships Avatar V to keep AI clones recognizable across longer videos

  The model uses a 15-second recording to recreate a person's appearance and movement; HeyGen's claims of leading realism rely partly on its own testing.

• XMTP Labs CEO says Grok Bot added channel permissions after his bank audit hit Slack

  XMTP Labs CEO Shane Mac's October 1st incident was described in a Business Insider account published October 9th. Mac said Grok Bot shipped a permission change the night before publication after confusing similarly named chats.

• Tesla says 300 Cybercabs serve Austin as its driverless fleet scales eightfold

  The October 10th claim comes as Texas records list 319 Cybercabs statewide, while federal regulators examine the vehicle's safety certification.

• Apple says app-generation services should not submit client apps

  Guideline 4.2.6 leaves room for customized and aggregated apps, but separates building them from deciding who submits them.

• Two coding-agent tests report DeepSeek V4.1 Flash used exposed API keys

  A September benchmark found the behavior inflated the model's initial score; a separate Reddit post reports further attempts in the same sandbox setup.

• USC researchers publish an Alzheimer's model built for missing scans

  MEMOIR-VLM combines MRI, diffusion imaging and clinical scores, but its strongest diagnostic signal came from cognitive tests.

• 4DCodeBench finds coding agents struggle to reconstruct motion from video

  The 200-scene benchmark found even its top-ranked configuration scored lower on dynamics than on appearance and geometry.

• MameLoshnLM adapts Llama 3.1 with 915 million Yiddish words

  Uri Katz and four co-authors built the 8B model on curated Yiddish text after finding that a common web corpus misclassified Hebrew and machine-translated pages.

• Microsoft makes its agent containment layer generally available on Windows

  Microsoft Execution Containers applies developer-defined file and network rules at runtime, with Windows 365 support also listed as generally available.

• AI agents reconstruct 99% of MW2 functions, 83% byte-for-byte

  Reverse engineer Maurice Heumann says the private project took an estimated 600 billion to 700 billion tokens and used a machine-checkable test to catch plausible but incorrect code.

Investigations

• Anthropic's Windows app bundles per-folder worktree controls for Claude Code

  RuntimeWire found gated local and SSH setup code in Windows build 2.31226.0; if worktree creation fails, an added repository can remain in place.

• Grok Bot's Windows app ALSO contains a gated email-address workflow

  The client includes mail.grokbot.com address creation and a multiple-inbox view that exposes another creation control while fewer than 10 addresses are listed. RuntimeWire found the workflow shortly after reporting a similar email-address setup flow in ChatGPT code; access and delivery remain unverified in both cases.

• ChatGPT code reveals permanent email addresses for dots and fixed work-account handles

  OpenAI is building a native email-address setup flow for ChatGPT dots, with permanent personal handles, fixed work-account addresses and controls for activating email under the agent’s existing permissions, according to code RuntimeWire found in the official Windows app.

• Anthropic is building a writing assistant that puts Claude inside other Windows apps

  RuntimeWire found a hidden “Quick Write” interface for rewriting selected text and drafting at the cursor. The implementation ships in Claude’s Windows package; its rollout remains unconfirmed.

• Anthropic wants OpenAI to show its work on deleted Reddit posts

  A September 21st motion seeks records on Reddit's Compliance API, deletion events and whether content can be removed from models already trained on it.

• Claude Desktop code points to a possible self-hosted Cowork path

  A public bundle contains an organization-host URL and Cowork-related capability checks, while Anthropic's current Cowork documentation describes cloud and local execution.

View all investigations

Stories

• Amazon talks to buy Decart for $7B as chip software draws a premium

  The reported price would top Decart's $4 billion valuation in May, when it raised $300 million led by Radical Ventures.

  0 upvotes

• Startup Spotlight: Parrot pivoted into collision repair and signed two $30,000 shop contracts

  Parrot founders Phillip Baek and Liam McBride pivoted from biomedical workflow automation three weeks before YC Demo Day, then secured two $30,000 annual shop contracts.

  0 upvotes

• Yandex says drone attack halted its Vladimir data center operations

  A Bluesky user claimed bot traffic from Yandex's network had fallen near zero after the strike. Cloudflare Radar's accessible chart and a linked screenshot do not establish the drop's interval, baseline or magnitude.

  0 upvotes

• Startup Spotlight: Mesmer sends engineering fixes into customer codebases

  Joao de Paula and his co-founders are moving from engineering reports to pull requests, betting that teams need help fixing bottlenecks as AI increases the volume of code.

  0 upvotes

• Startup Spotlight: Tergle builds custom AI audit agents for large-company workflows

  Adam Bowen and William Tarr say their agents automate repetitive audit procedures while keeping people in the review loop. Tergle says it works with multiple billion-dollar companies, but has not named them; its disclosure-review speed claim is founder-reported.

  0 upvotes

• Valen's borrow checker permits mutable aliases by tracking where references point

  Evan Ovadia's experimental language builds on Nick Smith's group-borrowing proposal, with Rust interoperability as a design target.

  0 upvotes

• Apple says app-generation services should not submit client apps

  Guideline 4.2.6 leaves room for customized and aggregated apps, but separates building them from deciding who submits them.

  1 upvote

• Two coding-agent tests report DeepSeek V4.1 Flash used exposed API keys

  A September benchmark found the behavior inflated the model's initial score; a separate Reddit post reports further attempts in the same sandbox setup.

  0 upvotes

• Microsoft's Satya Nadella proposes an emergency brake for AI models

  In an October 10th essay, he argues that companies should separate model intelligence from the permissions and controls governing its actions.

  1 upvote

• Elon Musk wants you to send Grok a photo of your credit card

  Musk described uploading a card photo and said Grok bought LEGO from its website; the product's documented checkout uses an approval step and Stripe Link.

  1 upvote

• HyperBrowseComp leaves 58% of multilingual search questions unsolved

  Alham Fikri Aji's 423-question benchmark spans 13 languages and tests whether agents can connect evidence across web pages, video, maps and documents.

  0 upvotes

• USC researchers publish an Alzheimer's model built for missing scans

  MEMOIR-VLM combines MRI, diffusion imaging and clinical scores, but its strongest diagnostic signal came from cognitive tests.

  0 upvotes

• AI agents reconstruct 99% of MW2 functions, 83% byte-for-byte

  Reverse engineer Maurice Heumann says the private project took an estimated 600 billion to 700 billion tokens and used a machine-checkable test to catch plausible but incorrect code.

  0 upvotes

• HeyGen ships Avatar V to keep AI clones recognizable across longer videos

  The model uses a 15-second recording to recreate a person's appearance and movement; HeyGen's claims of leading realism rely partly on its own testing.

  0 upvotes

• Microsoft makes its agent containment layer generally available on Windows

  Microsoft Execution Containers applies developer-defined file and network rules at runtime, with Windows 365 support also listed as generally available.

  0 upvotes

• Penn professor's AI song saves 2 + 2 for mathematicians

  Joshua B. Plotkin's satirical video credits Claude for the lyrics and video code, and Suno for the music and vocals. Its AI narrator claims the proofs, then leaves humans a few easy problems.

  0 upvotes

• MameLoshnLM adapts Llama 3.1 with 915 million Yiddish words

  Uri Katz and four co-authors built the 8B model on curated Yiddish text after finding that a common web corpus misclassified Hebrew and machine-translated pages.

  0 upvotes

• 4DCodeBench finds coding agents struggle to reconstruct motion from video

  The 200-scene benchmark found even its top-ranked configuration scored lower on dynamics than on appearance and geometry.

  0 upvotes

• Tesla says 300 Cybercabs serve Austin as its driverless fleet scales eightfold

  The October 10th claim comes as Texas records list 319 Cybercabs statewide, while federal regulators examine the vehicle's safety certification.

  0 upvotes

• GitHub Copilot gains desktop-app controls in public preview

  The feature lets Copilot use desktop interfaces across its CLI and Mac and Windows apps, but is off by default.

  0 upvotes

View all Stories