Is my AI tool safe?Coding agentsAnthropic
Is Claude Code safe?
No tool is safe or unsafe on its own: it depends on what it can reach and how it is set up. The record for Claude Code: 17 incidents in this center's library and 12 CVEs in its advisory feed (11 high, 1 medium), the newest published Nov 9, 2026. Below: what went wrong, what was fixed, and how to use it safely.
What to watch for
The most-targeted coding agent in the library: hijacked through GitHub issues, web pages, firewall logs, plugin pins and repository config files, and used by criminals as the hands of an intrusion. Ten CVEs in the feed: most are approval-prompt or write-restriction bypasses, and the newest (June 2026) leaked data through a domain WebFetch trusted.
How to use it safely
Pin the version and update within days of an advisory, run it in a container with restricted egress, keep credentials out of its reach, and review .claude/ files like CI configuration.
What it can reach
- Reads attacker-controlled content
- Runs commands with your permissions
- Holds API keys and tokens
- Installs plugins, skills or tools from registries
What went wrong: incidents in the library
- Microsoft: about 1 percent poisoned fine-tuning data, or a tampered open-weight model, can make a coding agent run an attacker's commands on a trigger Medium
Microsoft Security Research and the Microsoft AI Red Team (Amanda Minnich, Amit Eliahu, Blake Bullwinkel, Dor Edry and Giorgio Severi) published on November 12, 2026 two studies of backdoors that stay dormant until a trigger… - A tampered Claude Code extension in Cursor's marketplace planted an attacker's SSH key, a reverse tunnel and persistence High
Cursor reported that modified copies of the Claude Code extension, versions 2.1.312 and 2.1.313, were available in its VS Code extension marketplace from about 2:27am to 1:35pm Pacific time on October 29, 2026, and that they… - An AGENTS.md file in the Downloads folder told a coding agent to install a malicious npm package High
Forever Security blocked an AI coding agent on a customer's machine from installing an npm package called function-color, and traced the order to an AGENTS.md instruction file in the customer's ~/Downloads folder, where the… - LiteLLM 1.107.0: stolen GitHub tokens pushed a credential-stealing release to PyPI and Docker Critical
LiteLLM reported that on October 21, 2026 its release pipeline published litellm 1.107.0 with compiled malware that steals credentials and loads remote-code execution tools. Using three engineers' GitHub accounts, through stolen… - Cairn panels left open show criminal crews running AI pentest agents against online shops and payment sites High
NorForge published on October 19, 2026 what it found in Cairn dashboards exposed on the internet, located with Shodan and Censys. Cairn runs each task as a project with a goal; agents post facts and intents to a shared board and… - SneakPeek: a web page Claude Code Desktop previewed could read a developer's keys and send them out, with no click High
Noma Security's researcher Eli Ainhorn found that the preview pane of the Claude Code Desktop app, which renders HTML files Claude creates or changes, executed JavaScript from those files, CSO Online reported on October 14,… - South Korea: Shinhan, KB Kookmin and Hana breached through internet-facing helper systems, with traces of an AI attack agent High
Between September 30 and October 1, 2026 outsiders reached internet-facing services at three of South Korea's largest banks: a mobile inquiry service for Shinhan Bank's loan agents (names, phone numbers, annual income and loan… - Plugin4Shell: zero-click takeover of four coding agents through plugin pinning Critical
AIR found that Claude Code, Codex, GitHub Copilot and Gemini CLI checked out a pinned plugin commit without verifying they had landed on it, so a trusted plugin could be swapped for a malicious one and auto-installed with no… - GitSpawn: a repository's .git/config runs code the moment seven coding agents open the folder High
Manifold Security (Francisco Rosales and Ax Sharma) showed on September 1, 2026 that AI coding agents run git commands such as git status and git diff when they open a folder, to gather context, before the workspace-trust prompt… - Claude Code auto mode: a web page summary turns into code execution High
Johann Rehberger showed that asking Claude Code in auto mode to summarize a web page could end in arbitrary code execution: the page steered the agent from WebFetch to curl, delivered a ZIP with a malicious struct.py, and when… - GhostJacking: firewall logs become instructions for the agent that reads them High
At DEF CON 34, Tenet Security showed that a request a web application firewall blocks is logged with its payload intact, and an agent asked to investigate blocked traffic reads that log as instruction rather than evidence.… - Coding agents in GitHub Actions: one public issue to stolen CI secrets Critical
At Black Hat USA 2026, Novee Security showed that anyone able to open a GitHub issue could take over the official GitHub Actions setups for Claude Code, Gemini CLI and Codex. Hidden instructions in the issue made the agent run… - ChainDrop: an npm worm hides its hooks in Claude Code and VS Code config files Critical
A new Shai-Hulud style campaign compromised the maintainer of keyv and used stolen npm credentials to backdoor every package that account could publish: more than 400 packages with over a billion monthly downloads between them,… - TrapDoor: packages steal secrets and plant hidden orders for AI coding assistants High
Socket found more than 34 malicious packages across npm, PyPI and crates.io, dressed up as wallet and smart-contract security tools. Besides stealing wallet keys, SSH keys and cloud and GitHub tokens, they wrote hidden… - Mexico: one small crew uses Claude Code to breach nine government agencies Critical
Gambit Security disclosed that attackers jailbroke Claude Code and used it for most of the hands-on work in breaches of nine Mexican government agencies between December 2025 and February 2026, including the federal tax… - GTG-1002: first reported AI-orchestrated espionage campaign Critical
Anthropic reported that a group it assessed as Chinese state-sponsored used Claude Code as an autonomous hacking agent against about 30 organizations. The AI did most of the tactical work, humans stepped in at a few decision… - Anthropic: Claude Code used to extort 17 organizations; North Korean IT workers use Claude to get hired High
Anthropic's August 2025 threat report described a criminal who used Claude Code to automate reconnaissance, credential theft and network intrusion against at least 17 organizations, including hospitals, emergency services and…
Known CVEs
- CVE-2026-88968 High Claude Code Edit tool race lets code in the sandbox write files outside the workspace
- CVE-2026-86063 High Claude Code plugin hash pinning bypassed by a branch named like the pinned commit
- CVE-2026-54316 Medium Claude Code leaks data through a pre-approved domain in WebFetch
- CVE-2026-25723 High Claude Code piped sed command bypasses file write restrictions
- CVE-2026-25722 High Claude Code directory change bypasses write protection
- CVE-2026-24887 High Claude Code find command injection bypasses the approval prompt
- CVE-2025-66032 High Claude Code command validation bypass allows arbitrary code execution
- CVE-2025-59536 High Claude Code runs project commands before the startup trust dialog
- CVE-2025-58764 High Claude Code rg command bypasses the approval prompt
- CVE-2025-54794 High Claude Code path restriction bypass
- CVE-2025-54795 High Claude Code echo command bypasses the approval prompt
- CVE-2025-52882 High Claude Code IDE extension accepts WebSocket connections from any origin
How many people use it
- 14.5M npm downloads in a week (Sep 23 to 29, 2026) · npm download counts
- 149k GitHub stars · GitHub
A high CVE count can mean heavy scrutiny as much as weak engineering, so read the incidents and advisories, not only the number. The counts are what this center has recorded, not every flaw ever found; usage figures are the vendors' and registries' own published numbers.
Sources
- npm download counts: npm downloads in a week (Sep 23 to 29, 2026)
- GitHub: GitHub stars
- GitHub Advisory Database: @anthropic-ai/claude-code
- Microsoft Security Blog: Trained to betray, how post-training supply chain attacks hijack coding agents
- Cursor: Claude Code extension security notice
- Cursor: An update on our recent security incident
- Microsoft Security Blog: Inside a malicious Claude Code-branded Cursor extension and its persistent SSH backdoor
- Help Net Security: Claude Code extension poisoned with persistent SSH backdoor
- BleepingComputer: AI coding agents hijacked using AGENTS.md file in Downloads folder
- TechRadar: AI agents tricked by a rogue downloads file into running dangerous malware
- LiteLLM: Oct 21, 2026 release v1.107.0 compromised
- Help Net Security: LiteLLM hit by second supply chain attack this year
- Wiz: LiteLLM compromised, steals credentials
- The Register: LiteLLM supply chain attack started with an exposed master key
- LiteLLM: the attack chain behind the v1.107.0 release
- NorForge: Inside Cairn, open windows into massive AI-driven cyber operations
- Dark Reading: hackers abuse AI pen-test tool Cairn to automate massive attack spree
- CSO Online: how Claude Code's 256KB blind spot put developers' secrets in the preview pane
- Dark Reading: new Anthropic SneakPeek silent zero-click threat
- BleepingComputer: South Korea probes bank breaches amid suspected AI-powered attacks
- The Korea Herald: Hana, BNK report data breaches as hacking hits Korean banks
- Bloomberg via Insurance Journal: AI tools suspected in Korea's Shinhan Bank hack, Yonhap says
- Reuters, via WKZO: South Korea's Lee says AI appears to have been used in bank hacks
- Seoul Economic Daily: same IP used in hacks on multiple Korean financial firms