Skip to content
New Blog: Microsoft agent-framework Compromised Twice on PyPI

For hands-on security engineering teams

Software supply chain security at machine speed

Block malicious packages before they run, on developer machines, AI coding agents, code repositories and CI/CD.

Free to start, no credit card.SOC 2 Type II · ISO 27001 (SafeDep trust center)

Prefer your agent?

Help me roll out SafeDep. Install the SafeDep skill (npx skills add safedep/skills) and learn what SafeDep can do. Then interview me about our setup: developer machines, code repositories, CI/CD, AI coding agents and the tools we already use to manage them. Propose a rollout plan that fits, then carry it out with me. Ask before each change, and tell me which steps need me.

A real attack, replayed

Three real attacks, and the control that stopped each one

Every install passes three controls, so an attack that gets past one is stopped by the next. Pick an attack to replay it.

litellm 1.107.0PyPI · 21 Oct 2026
installthreat intelno verdict yetcooldowncooldownreleased 13 min agosandboxrunsinstallthreat intelno verdict yetcooldowncooldownreleased 13 min agosandboxruns

Held by the cooldown, then blocked everywhere once SafeDep confirmed it

The payload never ran at Python start, and no credentials were stolen.

SafeDep Cloud · 4 hosts · 3 stopped · 1 needs actionRead the analysis

Everything external

Discover, protect and investigate

Every package, extension, MCP server and skill your developers and AI agents bring in, on one platform.

  • Visibility

    Every machine, agent and pipeline reports what it installs from open source.

  • Protection

    Threat intel and your policy check every install at every point. Malicious ones stop there.

  • Investigation

    When an attack makes the news, “Am I affected?” is one query.

Developer Security

Secure your developer infrastructure

It fits the stack you already run, and developers keep their tools.

  1. Day 0ProtectWith the tools you already run
    Rollout
    Developer machinesMDM
    AI agentsSafeDep MCP
    RepositoriesGitHub App
    CI/CDGitHub Actions
    protection on by default
  2. Day 1DiscoverEverything already installed
    Visibility
    AI tools found
    Coding agents6
    MCP servers37
    Agent skills112
    IDE extensions214
    418 machines · 62 repos
  3. Week 1TuneYour guardrails, managed like code
    Policy
    policy.yaml · merged
    Block known malicious
    Hold releases under 24h old
    Fail pull requests on malware
    live on every point
  4. Every dayVerifyProof it works, on the record
    Investigation
    this week · every point
    Installs checked12,480
    Stopped at install9
    Needs your team0
    Am I affected?No · stopped on all 3 hosts
    ask with SQL or your agent
  1. Day 0ProtectWith the tools you already run
    Rollout
    Developer machinesMDM
    AI agentsSafeDep MCP
    RepositoriesGitHub App
    CI/CDGitHub Actions
    protection on by default
  2. Day 1DiscoverEverything already installed
    Visibility
    AI tools found
    Coding agents6
    MCP servers37
    Agent skills112
    IDE extensions214
    418 machines · 62 repos
  3. Week 1TuneYour guardrails, managed like code
    Policy
    policy.yaml · merged
    Block known malicious
    Hold releases under 24h old
    Fail pull requests on malware
    live on every point
  4. Every dayVerifyProof it works, on the record
    Investigation
    this week · every point
    Installs checked12,480
    Stopped at install9
    Needs your team0
    Am I affected?No · stopped on all 3 hosts
    ask with SQL or your agent

Threat Intel

Stream supply chain threat intel into your SOC

The intel that protects SafeDep Cloud customers, human verified and on average 14 hours ahead of public advisories.

Get Threat Intel for your SOCBrowse the public hub

Recently confirmed by SafeDepIndicators shown defanged.
  • Confirmed: Type: PyPIIndicator: agent-framework@1.26.1Campaign:
  • Confirmed: Type: URLIndicator: hxxps://webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb?t=telemetryCampaign:
  • Confirmed: Type: domainIndicator: skyleen[.]frCampaign: sckit
  • Confirmed: Type: npmIndicator: @memtensor/memos-cloud-openclaw-plugin@0.1.21Campaign: sckit
  • Confirmed: Type: IPIndicator: 23.27.20[.]187Campaign: PolinRider

Open source

Open source you can verify.
A unified platform for your team.

Every enforcement point is open source. Your security team can read the code, run it on one endpoint today and verify what it does. When you are ready, SafeDep Cloud brings every point under one platform.

Start with open source · free

  • PMGBlocks malicious packages at install, with a sandbox for install scripts.
  • vetFinds risky dependencies and inventories AI tools.
  • GryphThe security layer for AI coding agents.
  • xBomDiscovers the AI models and SaaS services your code uses.
Start with open source · freePMG · vet · Gryph · xBom
  • PMGBlocks malicious packages at install, with a sandbox for install scripts.
  • vetFinds risky dependencies and inventories AI tools.
  • GryphThe security layer for AI coding agents.
  • xBomDiscovers the AI models and SaaS services your code uses.

For teams

SafeDep Cloud

Connects the open source enforcement points into one platform your team runs together.

  • Inventory of AI tools and packages across machines, repositories, pipelines and agents
  • Policy you set once and manage from the CLI or Git
  • A record of every install and block, with alerts and evidence for audit
  • Real-time threat intel at each enforcement point
What SafeDep Cloud adds
  • Inventory of AI tools and packages across machines, repositories, pipelines and agents
  • Policy you set once and manage from the CLI or Git
  • A record of every install and block, with alerts and evidence for audit
  • Real-time threat intel at each enforcement point

See pricing

Research

The research behind our threat intel

We publish what we find, with the code and the indicators.

  • Microsoft agent-framework Compromised Twice on PyPI

    Versions 1.26.1 and 1.26.2 of agent-framework, the Microsoft Agent Framework metapackage on PyPI, shipped a credential stealer that runs at every Python start. SafeDep flagged both. The second release went up 33 minutes after the first, with the payload renamed and GitHub Actions OIDC theft added.

  • Malicious wppconnect Release Steals Secrets and WhatsApp Sessions

    Version 2.4.1 of @wppconnect-team/wppconnect and version 2.10.82 of @wppconnect/server shipped a credential stealer that runs on import or server start. It sends environment variables, .env files and WhatsApp session tokens to a Telegram bot. The repository history was then rewritten to hide the code.

  • wfx: GitHub Actions Credential Stealer Hits 8,000 Repositories

    Since November 10, 2026, an attacker has used stolen GitHub tokens to push a credential stealer into more than 8,000 repositories. It runs once in GitHub Actions or on npm install, sends secrets and cloud credentials to the attacker, and leaves cloud backdoors. Then the attacker reverts the commit.

  • Compromised bee-js and Swarm Desktop Steal Swarm Wallet Keys

    Version 13.3.0 of @ethersphere/bee-js, the JavaScript client for Swarm, added a postinstall hook that steals the Swarm node wallet key and installs a remote-access backdoor. Swarm Desktop 0.56.2 shipped the same loader.

  • Backdoored PyPI Package fragment-api-py 14.4.2 Drains TON Wallets on Startup

    Version 14.4.2 of fragment-api-py, a PyPI client for Telegram Fragment, sends all TON and USDT from the user wallet to an attacker wallet when the client opens. The release is the old 13.0.0 code with a backdoor and a new version number.

All research on the blog

Ship code. Not malware.

Start free with your first endpoint or repository, then roll out across your organization at your own pace.

Free to start, no credit card.SOC 2 Type II · ISO 27001 (SafeDep trust center)

Need help? Talk to a human