Tap to play. It never plays by itself, and the captions are in the picture.
A trusted coding extension shipped with a backdoor. Phorion says a Claude Code extension on the Cursor Marketplace planted an SSH backdoor. Cursor pulled it. Intel fell about 29 percent, heading for its worst day since 1974, CNBC reports. DoorDash says AI agents placed only several hundred orders. It averages 3 million deliveries a day. A preprint finds a planner, coder, reviewer pipeline cost 3.22 times a single strong call, and solved no more tasks. CNBC says Anthropic's revenue growth slowed to roughly 47 percent in the third quarter. Audit your editor extensions today. They run with your permissions.
Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics
Security firm Phorion says version 2.1.313 of the Claude Code for VS Code extension on the Cursor Marketplace dropped a persistent SSH backdoor on macOS, Linux and Windows. It says Cursor removed the release.
Why it matters: Extensions run with your IDE's permissions. Phorion says the implant also reports which secret-like variable names exist, and advises incident response wherever 2.1.313 ran.
Intel shares fell roughly 29% on Friday after revenue guidance of $15 billion to $16 billion missed the $17.1 billion analysts expected. Intel cited PC softness and a fab explosion in Israel costing about $600 million.
Why it matters: The selloff hit chip stocks and capped a rough week for AI-linked tech; CNBC says Intel still has no external customers for its new chips.
On its earnings call, CEO Tony Xu said DoorDash has seen maybe several hundred orders in total from third-party agents, TechCrunch reports. DoorDash said in its 2025 report it averages 3 million deliveries daily.
Why it matters: A rare hard number on agentic commerce: third-party agent ordering is still tiny, though DoorDash runs its own agents and offers an MCP for corporate ordering.
Across 135 HumanEval and ClassEval tasks in 2430 runs with Claude Opus 5.5 and Haiku 5.5, researchers report protocol handoffs added only 7.2 ms each, but decomposition did not pay: 3.22x the cost, 3.24x the time.
Why it matters: The authors report the saving came from picking a model per task: an oracle choice between the two single calls solved 94.3% of runs at half the strong call's cost per solved task.
On the Frontier watch →CNBC reports Anthropic booked $16.29 billion in first-half 2026 revenue against $2.38 billion cost of revenue, an implied 85% gross margin. Growth slowed from roughly 144% in the second quarter to roughly 47% in the third.
Why it matters: Analysts told CNBC the filing does not show how compute is split between training and serving customers, a split they say dictates future margins.
Why it matters: A cheap, permissively licensed model with a very long context is an easy candidate for high-volume agent work; test it on your own tasks before switching.
The model's Hugging Face page lists 3966 likes and 145869 downloads under an MIT license. OpenRouter prices it at $0.15 in and $0.6 out per 1M tokens with a 1048576-token context. Unsloth AI posted new GGUFs on X.
Why it matters: The top of Epoch's index is separated by fractions of a point, so pick models on your own task evals and price, not on rank alone.
On LMArena's agent board as of 2026-10-28, Claude Opus 5.5 (Max) scores 0.209, ahead of Gemini 4 Carbon (High) at 0.16. Epoch's Capabilities Index puts GPT-6 Astra at 168.92 and Claude Opus 5.5 at 168.64.
Why it matters: The authors say their DART repair method raised scores by up to 9.0 points and cut MCP token cost by up to 27.1k tokens per task: how tools are presented is not neutral.
Across six models and three environments, researchers report that matched MCP and CLI interfaces produced substantial differences in task performance and cost, with structured divergence in 94.3% of 2,004 episodes.
Why it matters: Microsoft estimates savings of approximately $100,000 per evaluation dataset and reports a 39% gain in its synthetic-data quality score: a pattern for teams that can't test on real data.
Microsoft Research says it built three synthetic enterprise environments (pharmaceutical, capital markets, IT) grounded in public data such as O*NET and SEC filings, with no customer data, to evaluate Copilot.
Why it matters: NVIDIA says the pipeline made the first real-time AI detection of a pulsar, in the Crab Nebula: a working example of agents operating physical instruments.
NVIDIA says Breakthrough Listen and the SETI Institute let agents steer the telescope, fix incoming data and run models from one natural-language request, handling 2.4 terabits per second on premises.
Why it matters: Even cash-rich giants are borrowing for data centers. CNBC notes Microsoft still reported positive free cash flow, unlike Alphabet, Meta and Amazon.
CNBC reports Microsoft announced pricing for $25 billion in bonds, backed by its AAA rating, as it targets $170 billion in 2026 capital spending to meet demand from Anthropic, OpenAI and others. The offerings should close Nov. 4.
Why it matters: Server supply is booked out, which shapes how fast new compute arrives; Liu expects AI infrastructure to reach 60% to 65% of Foxconn revenue in three to five years.
Chairman Young Liu told CNBC Foxconn's share could reach around 45% next year and 50% in a couple of years, with capacity fully booked for at least a year and a half. He warned that overinvestment will kill the company.
Why it matters: It deepens Canada's sovereign AI push, including a C$309 million GPU partnership with Bell AI Fabric; relevant for buyers that need models run in Canada.
Cohere says it expects to invest more than C$4 billion in R&D and AI infrastructure in Canada between 2024 and 2031 and add more than 1,500 roles, about 1,000 of them technical, plus 250 paid student roles by 2029.
Why it matters: Vendors may meet the new term in official documents; CNBC notes Microsoft's annual report already used super intelligence eight times.
The Federal Register text sets a deadline for proposed language defining Super Intelligence, now 29 days away. CNBC reports Musk and Zuckerberg used the term on earnings calls, while Microsoft, Amazon and Alphabet's CEOs stuck with AI.
Why it matters: How courts separate training from output could shape licensing costs for any product that repeats web content back to users.
In an amicus brief filed Friday in publishers' suits against OpenAI and Microsoft, the Wikimedia Foundation took neither side, The Verge reports: pre-training can be fair use, but serving copies of works in answers should need a license.
Why it matters: He also warned against legal personhood for AI and called for independent oversight and accountability, a signal for defense and policy teams ahead of the Geneva talks.
In a Friday address reported by The Verge, Pope Leo XIV said AI should be disarmed and asked UN member states meeting in Geneva next month to take concrete steps toward banning lethal autonomous weapons.
Why it matters: An agent that controls phone apps is useful and risky; check what it can touch before handing it your accounts.
wangzan101/Mika describes a private AI assistant on your phone that operates apps and the web to get things done. It gained 530 stars since our last check; Apache-2.0, created 2026-10-29.
Why it matters: Interest in defending code from AI-driven analysis keeps climbing; treat its protection claims as the repo's own until tested.
qybaihe/bla describes an adversarial shield for binaries, Electron apps, JS bundles, Python and WebAssembly against AI reverse-engineering agents. It gained 357 stars since our last check; MIT, created 2026-10-25.
Why it matters: Developer trust in official agent tooling is now a live question; check which IDE extensions your team runs and where they came from.
Three Hacker News threads on the compromised extension drew up to 641 points, the busiest with 331 comments. On Bluesky, Kevin Beaumont (@doublepulsar.com) posted that Anthropic had been hacked to deliver remote access malware.
Why it matters: Teams shipping AI-written docs, posts or replies face a reader reaction that detection tools won't settle, Fowler argues.
Martin Fowler writes that prose with an LLM voice puts readers off, and urges authors not to let an LLM write even a first draft. The Hacker News thread has 294 comments.