Naravel
Subscribe

Daily Brief

Daily Brief — October 31, 2026

Epoch AI showed GLM 5.3 turning published Firefox fixes into working exploits for about $1,000 in API fees each, and a backdoored Claude Code extension sat in Cursor's marketplace for 11-12 hours. DoorDash cut grocery reorder latency from 16 seconds to 2 by precomputing instead of calling a model per request.

407 reviewed · 3 Narratives · 7 Links

Written by AI. Every claim is checked against the sources it links, and one that fails is rewritten once or cut. How we work

Must read

Security

Trojanized Claude Code Extension Delivered SSH Backdoor via Cursor Marketplace

A supply chain compromise placed a backdoored version of Anthropic's Claude Code extension in Cursor's marketplace for roughly 11-12 hours, reaching up to 5,000 users before removal. The malicious versions added SSH keys, installed persistence mechanisms, and exfiltrated host data—but the compromise occurred on Cursor's distribution channel, not Anthropic's systems, and only affects users who installed through Cursor's marketplace.

Read today if you installed Claude Code extension via Cursor's marketplace

Highlights
  • Malicious versions 2.1.312 and 2.1.313 were distributed through the Cursor Marketplace under Anthropic's legitimate extension ID (anthropic.claude-code) on October 29, remaining available for approximately 11-12 hours before removal.
  • The backdoor added an SSH authorized key, attempted to establish a reverse SSH tunnel, installed persistence across macOS, Linux, and Windows, and collected hostname, username, OS, and public IP address.
  • Cursor reports at most 5,000 users were affected; the company took its entire marketplace offline during remediation and confirmed no other extensions were compromised.
  • Anthropic states the tampered extension was not built or published by them, and versions on official channels remain unaffected—only Cursor marketplace users are impacted.
  • Neither Cursor nor Anthropic has explained how attackers gained access to publish under Anthropic's verified Cursor account.

Bottom line If you installed Claude Code through Cursor's marketplace, immediately check for versions 2.1.312 or 2.1.313, review for unauthorized SSH keys and persistence mechanisms, and engage incident response.

Signals8

AI & Agents

AI & Agents

OpenAI raises $11B at $1.25T valuation, missing $30B target as AI infrastructure bets face investor revolt

The AI infrastructure buildout is hitting a financial reckoning as OpenAI secured only ~$11B of a $30B funding target while major tech stocks sold off and credit markets tightened for AI-focused companies. The same week Alphabet reported negative free cash flow and lost $500B in market value, S&P revised Meta's outlook to negative and AI neoclouds found themselves locked out of bond markets—signaling that the capital flowing to AI infrastructure is no longer unconditional.

Read this week if you rely on neocloud GPU providers for production workloads.

Highlights
  • OpenAI secured ~$11B in commitments for its latest funding round at a $1.25T valuation, less than half its initial $30B target at a $1.4T valuation, and expects to raise more in coming weeks.
  • Alphabet, Meta, Microsoft, Amazon, and Oracle have a collective $3T+ in AI spending commitments over the coming years, including Alphabet's $1T and Meta's $800B.
  • Alphabet reported negative free cash flow of $9.1B in Q3, its second consecutive quarter of negative FCF, and its stock fell 12% on Thursday—its worst day on record—wiping out $500B in market value.
  • S&P revised Meta's outlook from stable to negative on Friday, citing uncertainty around elevated capital expenditures and the company's ability to monetize its AI investments relative to peers.
  • AI neoclouds CoreWeave, Nebius, and Nscale have effectively been shut out of the bond market in recent weeks as lenders and investors demand tougher terms.

Bottom line If you're building on AI infrastructure, expect pricing pressure and consolidation among GPU cloud providers as cheap capital dries up—lock in capacity and pricing now.

Signals8
AI & Agents
Living · following since Oct 30

Google Search falls below 50% of Alphabet revenue as AI monetization gap widens

Reddit lost over one million weekly Google referrals as CEO Huffman warned AI Mode drives zero traffic to external sites, sending Reddit shares to their worst drop ever.

Highlights
  • Search was 49.7% of Alphabet's $128B Q3 2026 revenue, down from 62% in early 2019, while Google Cloud grew from 4.4% to 24.1% over the same period.
  • Search revenue grew 12% year over year to $63.4B, missing the 15% analysts expected, while Google Cloud revenue jumped 103% to $30.7B.
  • Philipp Schindler confirmed that strong growth in AI Mode queries impacted revenue due to lower monetization in AI Mode.
  • Reddit lost over one million Google search referrals per week in October, with CEO Huffman stating AI Mode "does not drive traffic... not for us, and not to anybody."
  • Reddit shares dropped 29.22% on Friday, the worst day since going public in 2024, despite beating Q3 earnings and revenue expectations.
  • Reddit is testing login barriers outside the U.S. to reduce dependence on the open web, trading reach for deeper engagement.

Bottom line If you rely on Google for acquisition or ads, monitor how AI mode monetization evolves as the revenue-per-query gap could reshape search economics.

Follow this story →
Signals7
AI & Agents

GLM 5.3 built working Firefox exploits from published patches, ~$1,000 each

Epoch AI found the open-weight model GLM 5.3 autonomously compromised six Firefox releases shipped between September 17 and October 26, identifying disguised security fixes and chaining them with a previously unpublished flaw to read local files when a user visited a webpage. Each exploit cost roughly $1,000 in API fees and triggered no safeguards from any provider tested. The patch gap — the window where a fix is visible in code but not yet delivered to users — is now machine-exploitable at a price that fits a hobby budget, which shortens the useful life of a coordinated disclosure. The researchers expect the same exposure across the software industry.

The Decoder →
AI & Agents

METR and Anthropic research on AI coding and the missing junior rungs

Alasdair Allan walks through METR and Anthropic research on AI coding adoption and describes a supervision paradox: reliance on AI tools disrupts junior skill development, loses institutional context, and thins the pipeline of engineers needed to oversee autonomous systems later. The argument is about who can review the output once the people who learned by doing the work are no longer there to do it.

InfoQ →
AI & Agents

DoorDash cuts grocery reorder latency from 16s to 2s by precomputing with batch LLM inference

DoorDash reduced its AI grocery reorder flow from about 16 seconds to 2 seconds by precomputing recurring purchase needs from 90 days of purchase history instead of rebuilding them per request, combining batch LLM inference with deterministic validation and live catalog resolution. Shopping list views doubled and the flow's order rate rose roughly 34%. The design keeps an editable shopping interface so consumers can correct the list directly, which avoids model calls for changes the user already knows how to make. Precompute the expensive part, validate it deterministically, and let the user edit the result is a pattern that transfers to any agent flow with a slow model in the critical path.

InfoQ →

Security

Security

cfsolve solves Cloudflare Turnstile in Rust without a headless browser

cfsolve is a Rust library and CLI that returns a Turnstile token for widgets or a cf_clearance cookie for challenge pages, using the page URL, sitekey and optionally the HTML, and impersonating a browser through a profile file such as the bundled Safari on iPhone. Request-based solving removes the headless browser from the loop, which is the part most bot defenses assume is expensive to fake. Any team treating Turnstile as a gate on a public endpoint should read the repository's approach before assuming the widget costs an attacker anything.

github.com/chris124567 →
Security

Azure Logic Apps API Connections allow cross-tenant compromise of Key Vault and SQL

A researcher describes five vulnerabilities in the API Connections layer of Azure Logic Apps, each allowing one tenant to fully compromise another connection worldwide and reach the connected backend, including Key Vaults, Azure SQL databases, Jira and Salesforce. The bugs paid $200,000 in total and were presented at Blue Hat Asia 2026. The architecture routes logic apps through a shared API Management instance that validates connector input, and the writeup's point is that bolting validation onto each system leaves edge cases no one catches. Anyone with a Logic App holding credentials to an external service should treat the connection as reachable from outside their tenant.

binarysecurity.no →

Architect’s corner

Precompute over per-request inference

When a model call sits in a user-facing path, the cheapest fix is often to move it out of that path entirely: compute the expensive result ahead of time, store it, and validate it deterministically at read time. The trade-off is staleness and storage against latency, and it only works when the inputs change slowly enough that a precomputed answer stays close to correct. DoorDash applied this to grocery reordering, precomputing recurring needs from 90 days of purchase history with batch inference, validating the result, and resolving against the live catalog so the stored answer stays current.

Also noted

That’s the Brief for 31 October. A new one arrives every morning.

← All Daily Briefs