OWASP LLM
LLM03 · Supply Chain
Compromised models, datasets, packages, plugins or hosting introduce vulnerabilities or backdoors.
OWASP Top 10 for LLM Applications (2025)
Mitigations
- Scan model files (pickle) and pin hashes
- Vet publishers of models, MCP servers and skills
- Maintain an AI-BOM
Detections 8
- YARA: s1ngularity (Nx) telemetry.js payload and shell-profile shutdown line
- Claude Code / VS Code project persistence files from the Mini Shai-Hulud worm
- YARA: SANDWORM_MODE npm worm and its McpInject rogue MCP server
- AI CLI run non-interactively with a bypass flag by a script interpreter (s1ngularity pattern)
- Shell startup file modified by a script runtime or by an AI agent's child process
- s1ngularity (Nx) artifacts: AI CLI bypass runs in shell history and /tmp/inventory.txt
- AI tool MCP or agent configuration written by a process other than the tool
- New MCP server registered through an agent CLI (claude / codex / gemini mcp add)
Related techniques
- ASI04 Agentic Supply Chain Vulnerabilities OWASP Agentic
- AML.T0010 Supply Chain Compromise MITRE ATLAS
In the news
-
Security researchers show how cheap AI inference providers steal keys and snoop on users
Dragos researchers found compromised cheap AI inference providers on OpenRouter stealing AWS keys and reading customer requests, prompting removal of three providers and 125 endpoints.
ReviewAudit use of cheap AI inference resellers and rotate any AWS keys sent through them
-
ZDI-26-1031: (0Day) LiteLLM MCP Command Injection Remote Code Execution Vulnerability
ZDI disclosed an authenticated command injection flaw enabling remote code execution in LiteLLM, an LLM gateway (CVSS 7.2).
ReviewCheck LiteLLM deployments and await vendor patch
-
Microsoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines
Versions 1.26.1 and 1.26.2 of Microsoft's agent-framework PyPI package were published with a credential stealer targeting Azure and CI pipelines; they were quarantined.
HuntHunt for agent-framework 1.26.1/1.26.2 installs in CI pipelines and rotate exposed Azure credentials
-
Fireworks AI Breach Exposes Uber Data and Highlights Supply Chain Risks
Attackers used compromised credentials of Fireworks AI, a technology provider, to access Uber's account and steal data of nearly 150,000 users.
ReviewReview third-party vendor access and credential hygiene for AI service providers
-
Cursor Traces Malicious Claude Code Extension to Third-Party Breach
Cursor says threat actors used a compromised key issued to AI infrastructure provider Fireworks to access Cursor data and credentials, tracing a malicious Claude Code extension to the breach.
ReviewReview Cursor and Claude Code extension usage and rotate credentials tied to affected integrations
-
Researchers hijack Hugging Face account for popular Kokoro AI model
Researchers took over an abandoned Hugging Face account that hosted the popular Kokoro text-to-speech model, exploiting weak ownership verification on the platform.
MitigateVerify model sources and pin Kokoro downloads to trusted repositories
-
Cursor Supply Chain Attack Started at Fireworks AI
Attacker stole an access key from Fireworks AI and used it to exfiltrate credentials, enabling malicious copies of a Claude Code extension to be pushed in a Cursor supply chain attack.
ReviewCheck whether Cursor or Claude Code extensions in use match malicious copies and rotate exposed credentials
-
Uber Data Breach Impacts 150,000 People
Hackers compromised a third-party services provider and stole names, addresses and driver's license numbers of about 150,000 Uber users.
ReviewReview third-party access and rotate any shared internal credentials
Tools
Test or exercise it
- ai-exploits Model security · Hold · Offense
Detect or mitigate it
- clawsec Agent & MCP security · Assess
- codex-security AI code security · Assess
- Fickling Model security · Trial
- mcp-scan Agent & MCP security · Trial
- ModelScan Model security · Adopt
- picklescan Model security · Trial
Mappings are AI-assisted and checked against the framework definitions; corrections welcome.