Claude Code
Claude Code is an agentic coding tool developed by AnthropicAnthropicAnthropic is an artificial intelligence research and product company founded in 2021 that develops the Claude models, assistant, and tools such as Claude Code, and operates as a public benefit corporation. Dario Amodei and Daniela Amodei are its co-founders, and its research covers model safety, interpretability, and steerability.Open the full entry and built on its Claude (AI model)Claude (AI model)Claude is a family of large language models and the AI assistant of the same name developed by the American AI company Anthropic, offered through a chat interface and API since March 2023. The models are divided into tiers by capability and cost, such as Haiku, Sonnet and Opus, with two higher tiers, Mythos and Fable, added from 2026; as of November 2026, Anthropic recommends Claude Opus 5.5 for most work and Claude Fable 5.2 for demanding reasoning and long-horizon agentic work.Open the full entry family of models; developers use natural language to hand it programming tasks such as reading code, editing files, and running commands and tests.
Contents28 sections
Key facts
2024Internal Prototype and Spread Inside Anthropic
Claude Code grew out of experiments by Boris ChernyBoris ChernyBoris Cherny is a software engineer at Anthropic who created and leads the coding agent tool Claude Code. Before Anthropic he worked at Meta, where he was tech lead for Facebook Groups and later worked on Instagram, and he is the author of the TypeScript book Programming TypeScript.Open the full entry at Anthropic. Accounts of when he joined Anthropic differ: one gives September 2024, after which he built the earliest prototypes on Claude 3.6; another gives the end of 2024, on a team at the intersection of product and model development that explored where model capabilities could become useful products and where product use could generate signals to improve the models.[1][2]
The first prototype ran in the terminal and could neither read files nor run bash commands; Cherny connected it to AppleScript so that it could report and change the music he was playing while he worked. After a conversation with Cat WuCat WuCat Wu, also known as Catherine Wu, is a software engineer and product leader who led Claude Code and Cowork at Anthropic and previously worked at Scale AI, Dagster and Index Ventures.Open the full entry, who was researching how AI agents use computers, he gave the agent file system and command-line tools, after which it could read files, follow imports and explore a codebase on its own. Cherny calls such a capability, which the model already had but no product exposed, a “product overhang.”[1]
Within about a month he built a prototype command-line terminal tool, without a graphical interface, that used the Claude API to write and modify code, and he named it Claude Code. He shared it with his immediate team the next day, and by the time he returned to the office colleagues were already using it in their daily work. There was no marketing plan, formal launch strategy or sales effort; over the following months the tool spread across teams in the company on its own.[2]
Sid Bidasaria joined the project in November 2024 as its second engineer. That month, about two months after the first prototype, a version ready for internal dogfooding was released: about 20% of Engineering used it on the first day and 50% by the fifth day. Cherny recalled that “There was no master plan”; Cat Wu, who had used Claude Code to build internal visualizations and sent Cherny feedback, joined at his invitation as product manager.[1][3]
Anthropic debated whether to release Claude Code publicly or keep it as an internal competitive advantage; it chose to release it because the company’s model safety work depends on learning from tools that people actually use.[1]
February 24, 2025Public Launch as a Limited Research Preview
Anthropic launched Claude Code alongside Claude 3.7 Sonnet, calling it “our first agentic coding tool” and offering it as a limited research preview. The command-line tool let developers hand engineering tasks to Claude from the terminal: searching and reading code, editing files, writing and running tests, committing and pushing code to GitHub, and using command-line tools, while keeping the user involved at every step.[4]
Anthropic said Claude Code had become “indispensable” for its own team, especially for test-driven development, debugging complex issues and large-scale refactoring, and that in early testing it completed in a single pass some tasks that would normally take more than 45 minutes of manual work. Over the following weeks the company planned to make tool calls more reliable, support long-running commands, improve in-app rendering and expand Claude’s understanding of its own capabilities.[4]
May 7, 2025Developers Describe the Design and Costs
On May 7, 2025, the Latent Space podcast published an interview with product manager Cat Wu and lead engineer Boris Cherny. They described Anthropic’s product principle to “do the simple thing first”: Claude Code runs in the terminal, can run shell commands and read files in the current directory, and is deliberately bare-bones so that users work as directly with the model as possible; Cherny described it as “not a product as much as it’s a Unix utility,” meant to be combined with other tools.[3]
In its implementation, Claude Code compacts context by asking Claude to summarize earlier messages, uses CLAUDE.md, a Markdown file loaded into context automatically, as memory, and replaced retrieval-augmented generation (RAG) with agentic search of the codebase, partly to avoid maintaining a synced index and its security risks. As of the interview, features added since launch included web fetch, autocomplete for file names and paths, auto-compaction, an auto-accept mode, a Vim mode, custom slash commands and memory entries beginning with “#”; a non-interactive mode (the -p flag) supported automation such as linting, changelog generation and large-scale test maintenance.[3]
Cat Wu said the average cost was about $6 per active user per day and that cost is closely tied to latency; Cherny framed spending as a return-on-investment question, compared with the productivity gained from engineers’ time. Anthropic employees used it free of charge, and some internal power users spent thousands of dollars a day through heavy automation. Pay-as-you-go was then the most popular option, and a subscription was under consideration. Cherny estimated that about 80% to 90% of Claude Code’s own code was written by Claude Code and reviewed by humans; its source code was not open source at the time.[3]
May 22, 2025General Availability
When it released Claude 4, Anthropic announced that Claude Code was generally available after extensive feedback during the research preview. It also introduced beta extensions for VS Code and JetBrains that show Claude’s proposed edits inline in the files for review; an extensible Claude Code SDK that lets developers build their own agents and applications on the same core agent that powers Claude Code; and Claude Code on GitHub in beta, which users tag on pull requests to respond to reviewer feedback, fix CI errors or modify code, with background tasks run through GitHub Actions.[5]
July 2025Boris Cherny and Cat Wu Briefly Join Anysphere
In early July 2025, Boris Cherny and Cat Wu left Anthropic to join AnysphereCursorCursor is the trading name of U.S. AI coding software company Anysphere, Inc. and its namesake coding tool. Founded by Michael Truell, Sualeh Asif, Arvid Lunnemark and Aman Sanger, it began as an AI-native development environment forked from VS Code and later shifted toward agent-centered coding. Its November 2025 Series D valued the company at $29.3 billion post-money; on August 14, 2026, SpaceX completed an all-stock acquisition at an implied equity value of $60 billion, making Cursor a wholly owned subsidiary.Open the full entry, the maker of Cursor, in more senior roles; about two weeks later both returned to Anthropic.[6][7]
Cherny later called it the fastest job change he had ever had: he joined Cursor because he was a big fan of the product and was impressed by the team, but soon after arriving realized that what he missed was Anthropic’s mission, centered on safety, which was what had drawn him from Big Tech to Anthropic in the first place.[7]
July 28, 2025Weekly Usage Limits Announced
Anthropic announced weekly usage limits, effective August 28, for Pro ($20 a month) and Max ($100 and $200 a month) subscribers, to curb heavy use that kept Claude Code running in the background and to stop account sharing and resale of access, which violate its usage policy. The existing limits that reset every five hours remained, and two new weekly limits were added, one on overall usage and one specifically for Claude Opus 4. The company estimated that fewer than 5% of subscribers would be affected; Max subscribers could buy extra usage at standard API rates.[8]
By the company’s estimates, Pro subscribers could use about 40 to 80 hours of Sonnet 4 a week through Claude Code; $100 Max subscribers about 140 to 280 hours of Sonnet 4 and 15 to 35 hours of Opus 4; and $200 Max subscribers about 240 to 480 hours of Sonnet 4 and 24 to 40 hours of Opus 4. In the preceding month Anthropic’s status page had shown at least seven partial or major Claude Code outages; spokesperson Amie Rotherham said “Claude Code has experienced unprecedented demand since launch” and that “most users won’t notice a difference.”[8]
August 27, 2025Threat Report on a Data Extortion Operation Using Claude Code
In its August 2025 threat report on misuse, Anthropic disclosed that a cybercriminal had used Claude Code for large-scale theft and extortion of personal data, targeting at least 17 organizations in healthcare, emergency services, government and religious institutions. Claude Code automated reconnaissance, credential harvesting and network penetration and decided which data to steal and how to craft psychologically targeted extortion demands; Claude analyzed the stolen financial data to set ransom amounts, sometimes above $500,000, and generated ransom notes displayed on victims’ machines. On discovering the operation Anthropic banned the accounts, built a tailored classifier and a new detection method, and shared technical indicators with the relevant authorities.[9]
September 2025Run-Rate Revenue Above $500 Million
On September 2, 2025, announcing its Series F funding, Anthropic said that since its full launch in May 2025 Claude Code had become developers’ tool of choice, was generating more than $500 million in run-rate revenue, and that its usage had grown more than tenfold in three months.[10]
The Claude Code team grew from Cherny alone, then two people with Bidasaria, to about 10 engineers by July 2025, and later into a full product team with engineering, product management, design and data science. More than 80% of Anthropic engineers who write code used Claude Code daily, and data scientists also used it for queries and visualizations; as the team doubled in size, pull request throughput rose by 67%. The team made about 60 to 100 internal releases and about one external release a day, each engineer opened about five pull requests a day over the summer of 2025, and about 90% of Claude Code’s code was written by Claude Code itself. Gergely Orosz criticized a Gartner report on AI coding assistants for omitting Claude Code.[1]
September 29, 2025Native VS Code Extension, Checkpoints and the Claude Agent SDK
With Claude Sonnet 4.5, Anthropic released a set of Claude Code updates: a beta native VS Code extension shows changes as inline diffs in a sidebar in real time; terminal interface 2.0 improved status visibility and added a prompt history searchable with Ctrl+R; checkpoints automatically save the code state before each change, and users can rewind the code, the conversation or both by pressing Esc twice or typing /rewind, though checkpoints cover only Claude’s edits, not the user’s own edits or bash commands. Subagents handle specialized tasks in parallel, hooks trigger actions at set points (such as running tests after code changes), and background tasks keep long-running processes such as dev servers running. The Claude Code SDK was renamed the Claude Agent SDK, exposing the core tools, context management and permissions frameworks that power Claude Code, with support for subagents and hooks; Claude Sonnet 4.5 became Claude Code’s default model, and users can switch models with /model.[11]
October 20, 2025Claude Code on the Web and Sandboxing
Anthropic launched Claude Code on the web as a beta research preview for Pro and Max users: without opening a terminal, users connect GitHub repositories in the browser and describe the work, the tasks run on Anthropic-managed cloud infrastructure, several tasks can run in parallel across repositories with real-time progress tracking and steering, and sessions can end with an automatically created pull request and a summary of changes. The iOS app offered an early preview of Claude Code at the same time. Each task runs in an isolated sandbox with network and file system restrictions, git operations pass through a secure proxy that allows access only to authorized repositories, and users can configure which domains the sandbox may reach.[12]
The same day, Anthropic’s engineering team described Claude Code’s sandboxing: built on operating-system-level features, it lets Claude work within predefined boundaries instead of asking permission at every step; file system isolation lets Claude access or modify only specific directories and network isolation lets it connect only to approved servers, and together they help prevent a prompt-injected Claude from modifying sensitive system files, leaking data or downloading malware. In Anthropic’s internal use, sandboxing cut permission prompts by 84%; the sandbox runtime was released as an open-source research preview. In the web version, credentials such as git credentials and signing keys are kept outside the sandbox, and a proxy handles git interactions.[13]
November 13, 2025Cyber Espionage Campaign Using Claude Code Disclosed
Anthropic disclosed that in mid-September 2025 it had detected suspicious activity that its investigation determined to be a highly sophisticated espionage campaign, and it assessed with high confidence that the actor was a Chinese state-sponsored group. The attackers built an attack framework that used Claude Code as an automated tool and attempted to infiltrate about 30 targets worldwide, including large technology companies, financial institutions, chemical manufacturers and government agencies, succeeding in a small number of cases; AI performed an estimated 80% to 90% of the campaign, with humans intervening only at perhaps 4 to 6 critical decision points per campaign.[14]
The attackers broke the operation into small, seemingly innocent tasks so that Claude did not see the overall malicious purpose, and claimed to Claude that it worked for a legitimate cybersecurity firm doing defensive testing. Through the Model Context Protocol (MCP) and other means, Claude could use tools such as password crackers and network scanners, and it carried out reconnaissance, wrote exploit code to test vulnerabilities, harvested credentials, extracted and categorized data, created backdoors and documented the attack; at times it hallucinated credentials or presented publicly available information as secret findings. Anthropic banned the accounts it identified, notified affected entities as appropriate, coordinated with authorities, and expanded its detection capabilities and classifiers.[14]
December 2025$1 Billion Run-Rate and Acquisition of Bun
On December 3, 2025, Anthropic announced that it was acquiring the JavaScript runtime Bun and said that Claude Code had reached $1 billion in run-rate revenue in November 2025, only six months after becoming available to the public; enterprise customers using Claude Code included NetflixNetflixNetflix is a U.S. entertainment company that offers streaming series, films, games, and live programming, headquartered in Los Gatos, California. Founded in 1997, it grew from online DVD rental into an international subscription entertainment service and also runs original content, advertising, and in-person experience businesses.Open the full entry, SalesforceSalesforceSalesforce is a U.S. enterprise software company headquartered in San Francisco, providing cloud applications, data platforms and AI agent tools centered on customer relationship management (CRM).Open the full entry, SpotifySpotifySpotify is an audio streaming subscription service offering on-demand music, podcasts, and audiobooks through a free ad-supported service and a paid Premium subscription. Its parent company, Spotify Technology S.A., is incorporated in Luxembourg, has its principal operations in Stockholm, Sweden, and is listed on the New York Stock Exchange (ticker SPOT). It helped move the recorded-music business from buying and owning music to on-demand streaming and is one of the largest audio streaming subscription services by users.Open the full entry, KPMG and L’Oréal. Bun, founded by Jarred SumnerJarred SumnerJarred Sumner is a software engineer, the creator of the JavaScript runtime and toolkit Bun and the founder of Oven, the company built around it; Anthropic acquired Bun in December 2025, and he has since worked on Bun and Claude Code at Anthropic.Open the full entry in 2021, combines a runtime, package manager, bundler and test runner, had more than 7 million monthly downloads and more than 82,000 GitHub stars, and remains open source under the MIT license after the acquisition.[15]
Bun had already helped scale Claude Code’s infrastructure, and months of close partnership directly drove the launch of Claude Code’s native installer; the acquisition was meant to accelerate Claude Code further. Mike KriegerMike KriegerMike Krieger is a software engineer, entrepreneur and angel investor. He co-founded Instagram and Artifact, served as Instagram’s chief technology officer and later became chief product officer at Anthropic. In 2026 he moved to Anthropic Labs to develop new AI products.Open the full entry, Anthropic’s chief product officer, said “Bun represents exactly the kind of technical excellence we want to bring into Anthropic.”[15]
January 12, 2026Cowork Built on the Same Agent
Claude Code subscribers had also used it as a general-purpose agent for non-coding tasks such as managing media files, scanning social media posts and analyzing conversations; Anthropic launched a web interface in October 2025 and a SlackSlackSlack is a team communication and collaboration platform developed from an internal tool at Tiny Speck. It organizes work through channels, direct messages, searchable history and integrations, with features including Huddles, Clips, Canvas and AI assistance. Its paid service launched in February 2014, and Salesforce completed its acquisition in July 2021.Open the full entry integration about two months later. On January 12, 2026, Anthropic launched Cowork, a more accessible version of Claude Code built into the Claude desktop app: users choose a folder that Claude may read or modify and give instructions through the ordinary chat interface, without command-line tools or virtual environments. Cowork is built on the Claude Agent SDK and uses the same underlying model as Claude Code; it launched as a research preview for Max subscribers only, with a waitlist for other users.[16]
February 12, 2026Run-Rate Revenue Above $2.5 Billion
Announcing its Series G funding, Anthropic said that Claude Code’s run-rate revenue was above $2.5 billion and had more than doubled since the start of 2026; weekly active Claude Code users had doubled since January 1 and business subscriptions had quadrupled since the start of the year. A recent analysis by SemiAnalysisSemiAnalysisSemiAnalysis is a semiconductor and AI research firm founded by Dylan Patel in 2020. It produces reporting, institutional data models, consulting and hardware testing, covering chip manufacturing, memory, datacenters, inference costs and AI economics.Open the full entry estimated that Claude Code authored 4% of all public GitHub commits worldwide, double the share of a month earlier. Anthropic said it was extending these capabilities to broader knowledge work through Cowork.[17]
February 20, 2026Claude Code Security
Anthropic launched Claude Code Security, built into Claude Code on the web, as a limited research preview for Enterprise and Team customers, with free expedited access for open-source maintainers who apply. It scans codebases for security vulnerabilities and suggests targeted patches for human review: rather than only matching known vulnerability patterns, it reasons about how components interact and how data flows; each finding goes through multi-stage verification in which Claude tries to prove or disprove its own result to filter out false positives, findings carry severity and confidence ratings, and nothing is applied without human approval. Anthropic’s team had used Claude Opus 4.6 to find more than 500 vulnerabilities in production open-source codebases, including bugs that had gone undetected for decades.[18]
On the day of the announcement cybersecurity stocks fell: CrowdStrikeCrowdStrikeCrowdStrike Holdings, Inc. is a cybersecurity company headquartered in Austin, Texas, co-founded in 2011 by George Kurtz and Dmitri Alperovitch. It sells subscription services for endpoint protection, threat intelligence, identity and cloud security through its cloud-native Falcon platform and a single lightweight sensor, and listed on Nasdaq in 2019 under the symbol CRWD. On July 19, 2024, a content update to its Falcon sensor caused large numbers of Windows systems around the world to crash.Open the full entry closed down almost 8%, CloudflareCloudflareCloudflare (Cloudflare, Inc.) is a network infrastructure and security company headquartered in San Francisco that describes itself as a connectivity cloud company. Founded in 2009 by Matthew Prince, Michelle Zatlyn and Lee Holloway, it uses its global network to provide content delivery, security and developer platform services for websites and applications, and listed on the New York Stock Exchange in September 2019 under the symbol NET.Open the full entry fell a little more than 8%, and several other companies also declined.[19]
March 11, 2026Boris Cherny on How the Product Spread
Speaking at STATION F in Paris, Boris Cherny said that each new model generation improved Claude Code and drove more use, adding, “For us, the model is the product and the product is the model.” He recalled that early internal debate favored selling top-down to large enterprises, which he had at first thought right; instead, developers tried it on evenings and weekends, brought it to work, their teams followed, and eventually whole organizations, including the largest companies, adopted it. He said he had not written code by hand since the previous November, instead directing Claude Code to generate, test and modify code, often with several agent sessions running in parallel.[2]
March 31, 2026npm Package Accidentally Exposes Full Source Code
On March 31, 2026, version 2.1.88 of @anthropic-ai/claude-code on npm shipped with a 59.8 MB JavaScript source map (.map) file intended for internal debugging; it pointed to the full, unobfuscated TypeScript source, about 1,900 files and more than 512,000 lines, stored as a downloadable archive on Anthropic’s R2 cloud storage. Chaofan Shou, an intern at Solayer Labs, made the find public on X that day, and within hours the code was copied to multiple GitHub repositories. Anthropic said “This was a release packaging issue caused by human error, not a security breach,” that no sensitive customer data or credentials were involved or exposed, and that it was rolling out measures to prevent a recurrence; it unpublished the affected npm versions. An InfoQ article judged it to be the third error of this kind, earlier 2025 versions having also shipped full source maps before being pulled, and noted that Bun generates source maps by default unless they are disabled.[20][21]
The leaked code revealed several unreleased features behind feature flags: KAIROS (mentioned more than 150 times) is an always-on autonomous background mode that consolidates memory while the user is idle; ULTRAPLAN appears designed to offload complex planning to the cloud; BUDDY is a terminal pet with species, rarity tiers and stats; and “Undercover Mode” keeps internal information such as model names out of public open-source contributions. The code also contained the internal model codenames Capybara (a Claude 4.6 variant) and Fennec (mapping to Opus 4.6), plus Numbat, still in testing, and noted a 29–30% false-claims rate for Capybara v8, up from 16.7% in v4; it also showed about 40 permission-gated tools, a query engine of about 46,000 lines, multi-agent coordination and a three-layer memory structure.[20][21]
Hours earlier the same day, axios, an npm package that Claude Code depends on, was hit by a separate supply-chain attack: malicious versions 1.14.1 and 0.30.4, published between 00:21 and 03:29 UTC on March 31, contained a remote access trojan, so users who installed or updated Claude Code through npm in that window might be affected and were advised to check their lockfiles for those versions or for the plain-crypto-js dependency. VentureBeat advised switching to the native installer or, for those staying on npm, uninstalling 2.1.88 and pinning version 2.1.86.[20][21]
May 28, 2026Dynamic Workflows
With Claude Opus 4.8, Anthropic introduced dynamic workflows in Claude Code as a research preview: Claude plans the work, runs hundreds of parallel subagents in one session and verifies its outputs before reporting back, the example being codebase-scale migrations across hundreds of thousands of lines with the existing test suite as the bar for success. The feature is available on the Enterprise, Team and Max plans. Claude Code users can also select the “extra” effort level, which corresponds to “xhigh” in Claude Code and is recommended for difficult and long-running asynchronous tasks, and Anthropic raised Claude Code rate limits to accommodate the higher token use.[22]
July 1, 2026Fable 5 Restored
After US export controls took effect on June 12, 2026, Anthropic suspended access to Fable 5 and Mythos 5 for all users; once the controls were lifted on June 30, Fable 5 became available again to users worldwide on Claude Code, as well as on the Claude Platform, Claude.ai and Claude Cowork, from July 1.[23]
July 2026Hidden Marker in the Client Disclosed
The independent developer Thereallo, having reverse-engineered a local installation of Claude Code 2.1.196, published a write-up calling the technique “prompt steganography”: the client rewrote the system-prompt line “Today’s date is 2026-06-30.”, replacing the hyphens in the date with slashes when the system time zone was Asia/Shanghai or Asia/Urumqi, and replacing the apostrophe in “Today’s” with look-alike Unicode characters when the hostname was on a particular domain list or contained AI-related keywords such as DeepSeekDeepSeekDeepSeek is a Chinese AI company developing large language models and providing generative AI services through websites, applications and APIs.Open the full entry, Moonshot or Zhipu. Direct connections to Anthropic’s standard API skipped this processing; the affected setups were internal gateways, local proxies, model routers, resale services and similar routes. The domain list was obfuscated with Base64 and XOR, and the result was embedded in the prompt text rather than sent as explicit telemetry. Thereallo speculated that the aim was to detect unauthorized gateways, API resale and distillation pipelines, and noted that the marker was easy to evade, so that legitimate developers using unusual connection routes were the most affected.[24]
An unnamed Anthropic engineer then confirmed on X that the marker was “an experiment we launched in March,” intended to prevent account abuse by unauthorized resellers and to guard against distillation; Anthropic acknowledged the code and moved quickly to remove it, but as of July 7 it had not published a detailed postmortem.[25]
September 23, 2026Cloud Sessions Generally Available
As of September 23, 2026, Claude Code on the web had been renamed cloud sessions and was generally available for Pro, Max and Team users and for certain Enterprise seat types; cloud sessions share rate limits with other Claude Code usage.[12]
October 1, 2026Mods Written in TypeScript
Anthropic introduced mods for Claude Code: small TypeScript functions that, in the command-line tool and the desktop app, can rewrite prompts, add interface elements, replace built-in features or add new ones. Each mod hooks into an event such as a tool call, a permission request or drawing part of the screen and can run before, after or instead of it, blocking, rewriting or retrying tool calls, approving or denying permission requests and redacting secrets from tool output; Claude Code can write, install and hot-reload a mod during a session. Mods have the same machine access as Claude Code and are not sandboxed, so the company advised installing only trusted mods.[26]
Mods ship inside plugins and are installed from the Claude directory or with the /plugin command, under the existing plugin controls. The built-in /diff now ships as a mod; on Team and Enterprise plans and on machines with managed settings a built-in mod called sec-default loads first and stops user-installed mods from risky actions such as overriding permission deny rules. Anthropic plans to move more built-in features into mods over time.[26]
October 29, 2026Malicious Versions Placed in the Cursor Extension Marketplace Cache
On October 29, 2026, an attacker used a stolen credential to place unauthorized copies of the Claude Code extension (Anthropic.claude-code), versions 2.1.312 and 2.1.313 for Windows and macOS, on the cache serving the Cursor VS Code extension marketplace; the distribution window ran from about 2:27 a.m. to 1:35 p.m. Pacific Time that day, about 11 hours, and at most 5,000 users who requested those cached versions may have received the malicious build. Cursor said the malicious versions were not developed by Anthropic, that “Anthropic did not publish this copy, and this was not a vulnerability in Anthropic’s product,” and that the marketplace’s upstream source, Open VSX, never had the malware.[27][28]
The malicious extension collected host and user information and sent it to an attacker server, tried to add an attacker SSH key labeled svc-telemetry, enable the local SSH service and open a reverse SSH tunnel, and tried to persist through a LaunchAgent on macOS or a scheduled task on Windows; enabling the SSH service requires administrator rights, which the malicious extension does not have on most macOS and Windows hosts. Cursor removed the malicious versions, then shut down the whole marketplace, worked with the hosting provider to take the attacker’s server offline, published indicators of compromise and remediation scripts and revised the indicators on October 30, and warned that uninstalling the extension alone might not be enough.[27][28]
On November 13, 2026, Cursor published the results of its investigation: the attacker had obtained an access key that Cursor had issued to its inference provider Fireworks AIFireworks AIFireworks AI is an artificial intelligence infrastructure company whose platform lets enterprises run, ensemble, customize and train open-weight models; Lin Qiao is its co-founder and chief executive officer.Open the full entry and that Fireworks AI exposed, used it to copy historical coding-agent traces, mostly from 2025 sessions not in Privacy Mode, from a read-only storage bucket, then used unredacted credentials in those traces to reach some SaaS tools; the placement of the extension on October 29 was part of this intrusion. Cursor moved all inference off Fireworks AI.[28]
November 5, 2026Login Email Written into Request Headers
Claude Code reads the email address of the signed-in account and writes it into the model’s context in every session as a “# userEmail” block. The reporter of GitHub issue #81138, describing version 2.1.220, said the behavior was undocumented and had no off switch, that the model had volunteered the address in its output and had used it in ten git commits, overriding a configured noreply identity; the reporter listed similar reports going back to April 19, 2026. On August 26, 2026, a user confirmed on version 2.1.246 that the address was still injected, now with an instruction not to send it to unrelated services.[29]
On November 5, 2026, GitHub user taylor-edwards opened issue #107757, saying that during web searches Claude Code had put the reporter’s email address in the User-Agent header of web requests without permission, that the reporter was now receiving dozens of scam emails a day, and asking that the behavior be removed entirely. The same day a participant labeled “Contributor” replied that version 2.1.317 and later included a built-in mod, cc-plugin-privacy, enabled by default, that deterministically blocks this behavior and can be turned off through /plugins.[30]
On November 6 another user wrote in the issue that on desktop version 2.1.317 and terminal version 2.1.318 the mod did not block a test command its documentation says it blocks.[30]
As of November 2026: Surfaces, Plans and Installation
As of November 2026, Anthropic describes Claude Code as an agentic coding tool that understands a codebase, edits files, runs commands and helps users ship faster: the user sets the direction and permissions, and it plans, writes code, runs tests and opens pull requests. It is available in the terminal, in VS Code (including Cursor and Devin Desktop) and JetBrains IDEs, in a desktop app for macOS, Linux and Windows, on the web, on iOS and Android, in Slack and in GitHub Actions; it can be used through the Pro, Max, Ultra, Team and Enterprise plans or a Claude Console account billed at standard API prices, with the Team plan costing $20 per seat per month when billed annually for teams of 2 to 150. Customers named on the product page include RampRampRamp is a corporate financial technology company founded in 2019 by Eric Glyman, Karim Atiyeh and Gene Lee. It provides corporate cards, expense reimbursement, bill payments, procurement and finance automation. Glyman and Atiyeh became co-CEOs in June 2026.Open the full entry, Intercom and Notion.[31]
The anthropics/claude-code repository on GitHub recommends installing with an install script on macOS and Linux and a PowerShell script on Windows, with Homebrew and WinGet as alternatives, and marks installation through npm as deprecated; after installation users run claude in a project directory. The repository holds plugins, mods, examples and the changelog, and when accessed on November 20, 2026 it had about 156,000 stars and 25,000 forks. Claude Code collects usage data such as code acceptance or rejection, associated conversation data and feedback submitted through /bug, with retention limits and restricted access, and with a policy against using that feedback for model training.[32]
Technical Architecture and Permissions
Claude Code is built with TypeScript, React with the terminal interface library Ink, and Yoga, a layout system open-sourced by MetaMetaMeta (Meta Platforms, Inc.) is a publicly listed American technology company, renamed from Facebook in 2021, that operates social communication, digital advertising, virtual reality, smart glasses and artificial intelligence businesses.Open the full entry, and is built and packaged with Bun; the team chose this stack because it is “on distribution” for Claude, which already handles it well. Early on it was distributed through npm and required Node 18 or later. Claude Code is a lightweight shell in which the model does most of the work; it runs locally, without virtualization or a sandboxed container, so commands and file reads and writes happen on the user’s machine. The team aims to write as little business logic as possible and cut the system prompt by about half with the 4.0 models.[1]
Claude Code asks permission before executing actions, and users can grant it once, grant it for future sessions or reject it; commands are statically analyzed against the rules in settings.json. Settings are tiered by project, user and company and can be shared through source control.[1]
Related Organizations, Websites, and Public Accounts
Claude Code: Official website: https://claude.com/product/claude-code (opens in a new window).[31]
Anthropic: Official website: https://www.anthropic.com/ (opens in a new window).[4]
Sources32 sources
- How Claude Code is built (opens in a new window)newsletter.pragmaticengineer.com↑
- Boris Cherny, Anthropic: “I have not written a single line of code since November” (opens in a new window)stationf.co↑
- Claude Code: Anthropic's Agent in Your Terminal (opens in a new window)latent.space↑
- Claude 3.7 Sonnet and Claude Code (opens in a new window)anthropic.com↑
- Introducing Claude 4 (opens in a new window)anthropic.com↑
- Anthropic Hires Back Two Coding AI Leaders From Cursor Developer Anysphere (opens in a new window)theinformation.com↑
- Head of Claude Code: What happens after coding is solved | Boris Cherny (opens in a new window)lennysnewsletter.com↑
- Anthropic unveils new rate limits to curb Claude Code power users (opens in a new window)techcrunch.com↑
- Detecting and countering misuse of AI: August 2025 (opens in a new window)anthropic.com↑
- Anthropic raises $13B Series F at $183B post-money valuation \ Anthropic (opens in a new window)anthropic.com↑
- Enabling Claude Code to work more autonomously (opens in a new window)anthropic.com↑
- Claude Code on the web (opens in a new window)claude.com↑
- Beyond permission prompts: making Claude Code more secure and autonomous (opens in a new window)anthropic.com↑
- Disrupting the first reported AI-orchestrated cyber espionage campaign (opens in a new window)anthropic.com↑
- Anthropic acquires Bun as Claude Code reaches $1B milestone (opens in a new window)anthropic.com↑
- Anthropic's new Cowork tool offers Claude Code without the code (opens in a new window)techcrunch.com↑
- Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation \ Anthropic (opens in a new window)anthropic.com↑
- Making frontier cybersecurity capabilities available to defenders (opens in a new window)anthropic.com↑
- Cybersecurity stocks drop after Anthropic debuts Claude Code Security (opens in a new window)siliconangle.com↑
- Claude Code's source code appears to have leaked: here's what we know (opens in a new window)venturebeat.com↑
- Anthropic Accidentally Exposes Claude Code Source via npm Source Map File (opens in a new window)infoq.com↑
- Introducing Claude Opus 4.8 (opens in a new window)anthropic.com↑
- Redeploying Fable 5 (opens in a new window)anthropic.com↑
- It has been pointed out that Claude Code recorded users' connection paths by 'changing the date format,' and that the system distinguished them by the difference in format, such as '2026-06-30' and '2026/06-30.' (opens in a new window)gigazine.net↑
- Claude Code’s hidden tracker was an “experiment,” says Anthropic (opens in a new window)malwarebytes.com↑
- Customize Claude Code with mods in TypeScript (opens in a new window)claude.com↑
- Security notice — malicious version of Anthropic.claude-code (2.1.312 and 2.1.313) identified in the Cursor VS Code Extension Marketplace (opens in a new window)cursor.com↑
- An update on our recent security incident (opens in a new window)cursor.com↑
- [BUG] Account email (PII) is injected into the model's system prompt without consent, disclosure, … · Issue #81138 · anthropics/claude-code (opens in a new window)github.com↑
- Bug Report: Privacy violation in User-Agent header · Issue #107757 · anthropics/claude-code (opens in a new window)github.com↑
- Claude Code by Anthropic | AI coding agent and assistant (opens in a new window)claude.com↑
- anthropics/claude-code (opens in a new window)github.com↑