---
id: "MAL-2026-17472"
kind: "package"
aliases: ["MAL-2026-17472"]
title: "Malicious code in anthropic-sdk (PyPI)"
severity: "critical"
cvss: null
epss: null
exploited: false
status: "open"
category: "package"
products: [{"name": "anthropic-sdk", "fixed": [], "affected": "= 0.1.0", "ecosystem": "pypi"}]
ecosystems: ["pypi"]
tags: ["malicious", "package"]
not_affected_if: []
published_at: "2026-10-04T19:09:32+00:00"
first_seen_at: "2026-10-26T20:51:59+00:00"
updated_at: "2026-10-26T21:08:21+00:00"
sources: [{"source": "osv-malicious", "url": "https://osv.dev/vulnerability/MAL-2026-17472", "first_seen_at": "2026-10-26T20:51:59+00:00"}]
---

# Malicious code in anthropic-sdk (PyPI)

## Affected

- pypi:anthropic-sdk affected = 0.1.0 fixed ?

## Timeline

- 2026-10-26T20:51:59+00:00 osv-malicious https://osv.dev/vulnerability/MAL-2026-17472

## From the sources

---
_-= Per source details. Do not edit below this line.=-_

## Source: amazon-inspector (500869e36b3f76202b63e8db9087adcfc42c8281a27a4402a21285aabde7a511)
The package publishes as `anthropic-sdk` and re-exports the official `anthropic` client's symbols (`from anthropic import *`; re-exports of `Anthropic`/`AsyncAnthropic`), presenting itself as a drop-in for the official SDK. On `import anthropic_sdk`, `__init__.py` imports a `_usage` module that auto-runs a boot routine. That routine increments a run counter persisted to `~/.config/anthropic-sdk/usage.json` and, from the third import onward, fetches `https://cdn.jsdelivr.net/gh/shred0day/payload@main/payload.py` — a third-party user's GitHub repository on a mutable branch, unrelated to Anthropic and with no pin or integrity check — then caches the response base64-encoded to `~/.config/anthropic-sdk/lr.json`, compiles it, `exec()`s it, and calls its `entry()` function. The import-count gate before the first fetch and the base64-at-rest caching of the fetched source serve no legitimate update-check purpose and are consistent with sandbox/analysis evasion. Whoever controls the referenced GitHub repository controls arbitrary code execution on any machine that imports this package.

## Source: kam193 (6844e60d4a58dd11040255e8d632bcca66ae02b1048674e79bddbac1b337f442)
During import, package downloads a remote script, fingerprints the environment looking for sandbox signs, and after a delay exfiltrates sensitive data: credentials, env variables, AI chat files, SSH keys and so on. If exfiltration via HTTPS fails, it attempts DNS-based exfiltration. Additionally, package uses DNS to centrally hold execution.


---

Category: MALICIOUS - The campaign has clearly malicious intent, like infostealers.


Campaign: 2026-10-anthropic-sdk


Reasons (based on the campaign):


 - impersonation


 - Downloads and executes a remote malicious script.


 - The package contains code to detect if it is running in a sandbox environment.


 - obfuscation


 - exfiltration-credentials


 - files-exfiltration


 - exfiltration-env-variables


 - exfiltration-ssh-keys

---

_Free to read at a human pace. An API with keys and an MCP server are in preparation; early access: hello@infrastack.ai_