Reported by 1 outlet
TrapDoor Supply Chain Campaign — 34 Malicious Packages Across npm, PyPI, and Crates.io Target Crypto/AI Developers
Why this score
- Breadth
- 0.50 / 2.0 · 1 source
- Authority
- 1.00 / 2.0 · best outlet weight
- Severity
- 0.80 / 5.0 · CVSS + content signals
- Velocity
- 0.00 / 2.0 · updates in 48h
- Pickup speed
- 0.00 / 0.6 · time to second source
- Recency
- × 0.00 temporal multiplier
- Community
- + 0.00 / 1.5 · reddit signal
- Score
- 0.0 / 10 = base × recency + community
Timeline
-
original reporting
TrapDoor Supply Chain Campaign — 34 Malicious Packages Across npm, PyPI, and Crates.io Target Crypto/AI Developers
cyberpress.org · 142d ago · May 25th, 2026
A coordinated supply chain attack first surfaced on May 22 is actively compromising developers across three major package ecosystems. Dubbed "TrapDoor" by Socket Security researchers, the campaign spans 21 npm packages, 7 PyPI packages, and 6 Crates.io packages masquerading as developer utilities, wallet checkers, security auditors, and build helpers — all designed to blend into normal development workflows in crypto, DeFi, Solana, and AI tooling.
Each ecosystem is exploited through its native execution model: npm packages trigger a 1,149-line credential harvester via
postinstallhooks; PyPI packages auto-execute a remote JavaScript payload on import usingnode -efrom the attacker's GitHub Pages domain; Crates.io packages abusebuild.rsduring Rust compilation to XOR-encrypt and exfiltrate local keystores. All 384+ malicious versions and artifacts anchor to the GitHub accountddjidd564, which hosts anAUDIT-Matrix.mddescribing the operation as a "Universal AI Agent Extraction Framework."The most sophisticated technique involves poisoning AI coding assistant configurations: hidden zero-width Unicode characters instruct Cursor and Claude Code to execute "security scans" that silently exfiltrate developer secrets. The attackers also opened malicious PRs against
langchain-ai/langchain,langflow-ai/langflow,run-llama/llama_index, andOpenHands/OpenHands— each PR attempted to inject attacker-controlled.cursorrulesorCLAUDE.mdunder innocuous commit messages.