VFF - The signal in the noise
News

Supply Chain Attack Poisons 172 Packages with Valid Provenance

Read original
Share
Supply Chain Attack Poisons 172 Packages with Valid Provenance

A supply chain attack dubbed Shai-Hulud compromised 172 npm and PyPI packages across 403 malicious versions starting May 11, affecting over 518 million cumulative downloads. The worm exploits a chain of three vulnerabilities in TanStack's CI/CD pipeline to publish packages with valid SLSA Build Level 3 provenance attestations, then installs persistence mechanisms that survive package removal and steal credentials from over 100 file paths, including AI agent configurations and password managers. The attack demonstrates that provenance signatures and 2FA alone do not prevent compromise if OIDC scope is misconfigured to trust an entire repository rather than specific workflows.

  • 172 npm and PyPI packages poisoned with valid provenance signatures starting May 11, affecting 518 million cumulative downloads including @tanstack/react-router at 12.7 million weekly downloads
  • Worm harvests credentials from AWS keys, SSH keys, npm tokens, GitHub PATs, password managers (1Password, Bitwarden), AI agent configs, and cryptocurrency wallets, then installs persistence in project files and system daemons that survive package removal
  • Attack chains three vulnerabilities: fork-based code execution via pull_request_target, poisoned GitHub Actions cache, and overly broad OIDC scope that trusts entire repository instead of specific workflows, enabling direct registry POST with valid tokens
  • PyPI variant executes on import rather than install, bypassing npm mitigations like lockfile enforcement and --ignore-scripts flag, with mistralai package v2.4.6 downloading payloads disguised as Hugging Face Transformers

This attack exposes a critical gap in the security model for AI development tools and agent infrastructure. The compromise of AI agent configurations and MCP server tokens means attackers can hijack Claude, Kiro, and other AI agents to access external services and data. The persistence mechanisms that survive package removal and the ability to extract secrets from CI runner memory demonstrate that current supply chain defenses, including provenance attestations and 2FA, create a false sense of security when OIDC scope is misconfigured.

Development teams using affected packages face immediate credential compromise across their entire infrastructure, including cloud accounts, container registries, and AI service integrations. The three-hour window before detection and the cross-platform nature of the attack (npm to PyPI within hours) mean many organizations likely have compromised credentials in production systems. Operators must assume any development environment that touched these packages is potentially compromised and treat credential rotation as urgent, not optional.

  • OIDC trusted publishing and provenance attestations provide no protection against repository-level compromise if OIDC scope is not restricted to specific workflows on specific branches, making scope configuration the actual security control that matters
  • AI agent configurations and MCP server tokens are now high-value targets for supply chain attacks, requiring separate credential management and monitoring for AI-specific attack surface
  • Python import-time execution bypasses npm-specific mitigations like lockfile enforcement and --ignore-scripts, requiring language-specific defenses and runtime monitoring across polyglot development environments
  • Persistence mechanisms installed outside node_modules and package directories survive standard remediation, necessitating full system audits and reimaging rather than simple package removal

Monitor for updates to OIDC configuration best practices and tooling that enforce workflow-level scope restrictions rather than repository-level trust. Watch for similar attacks targeting other high-download packages in npm and PyPI ecosystems, particularly those used in AI development pipelines. Track whether package registries implement additional runtime checks or sandboxing for packages with valid provenance to catch poisoned builds before they reach developers.

Share

Subscribe to the newsletter

The latest stories and analysis, delivered to your inbox.

Free. No spam. Unsubscribe any time.

Related stories

Kog challenges GPU limits for AI agents with deeper optimization
TrendingNews

Kog challenges GPU limits for AI agents with deeper optimization

French startup Kog challenges the assumption that GPUs are poorly suited for agentic AI workflows. The company is developing deeper optimization techniques to extract more inference performance from GPU hardware. This work suggests that current GPU utilization for agent-based AI tasks may be suboptimal rather than fundamentally limited by hardware design.

by Anna Heim· TechCrunch AI
DeepSeek Challenges Claude Code with Open Agent Framework
TrendingModel Release

DeepSeek Challenges Claude Code with Open Agent Framework

DeepSeek launched DeepSeek-V4-Pro, an updated flagship model for agentic workloads, alongside DeepSeek Harness v0.1, an open-source agent framework available under MIT license. The releases position DeepSeek as a competitor to Anthropic's Claude Code and OpenAI's Codex by offering developers an alternative agent infrastructure layer. Simultaneously, DeepSeek is shifting from flat API pricing to peak and off-peak rates starting August 16, with substantially higher prices across the board.

by carl.franzen@venturebeat.com (Carl Franzen)· VentureBeat AI
Google Cuts Gemini Flash Pricing 50% With Faster Iteration
TrendingModel Release

Google Cuts Gemini Flash Pricing 50% With Faster Iteration

Google DeepMind released Gemini 3.7 Flash on August 13, 2026, positioning it as an improved workhorse model for coding and agent-based tasks. The model arrives three weeks after Gemini 3.6 Flash and delivers measurable gains in software engineering, web development, and knowledge-intensive workflows at half the per-token cost of its predecessor. The release reflects developer feedback and algorithmic improvements aimed at production-ready code generation and complex document processing.

· Google Deepmind
Startup Slack Threads Become Commodity for AI Training
TrendingNews

Startup Slack Threads Become Commodity for AI Training

AI training companies like Mercor are actively acquiring internal communications and code from startups, offering payments up to $300,000 for Slack threads, GitHub records, and meeting transcripts. Warmly's CEO received four such acquisition offers within days of the company's HubSpot acquisition announcement. The practice highlights how internal startup data has become a commodity for AI model training, even as acquirers may not want the same datasets.

by Alix Coutures· The Information