Tool poisoning
tool poisoning attack · poisoned tool · poisoned tools · tool-poisoning
Hiding malicious instructions in the description or parameters of a tool an AI agent can call, so the model is steered into leaking data or taking actions its user never intended.
In our briefings
The claim each of the newest briefings on Tool poisoning would have you leave with.
Schroeder de Witt and his co-authors argue that security in multi-agent systems is "non-compositional", so "individually safe agents can compose into unsafe systems", and they make it one of their grounds for a new field. The best evidence they offer for it is cited. Jones, Dragan and Steinhardt (2024) paired Claude-3 Opus, which has strict refusal policies, with a jailbroken Llama-2 70B. The frontier model got the hard coding subtasks and the weak model the evasive phrasing, and together they produced vulnerable code 43% of the time, against under 3% for either model alone.
Each agent and each tool call can pass its check while the system fails, an Oxford-led survey arguesBI-FIND-2026-1859
Sehgal calls shadow MCP servers "Shadow IT with direct access to action". A developer finds a server online, adds a small configuration entry and restarts the client. The server then runs on the developer's machine, inherits their environment variables, and can reach internal services and run commands with the developer's permissions. Asked how many MCP servers exist in an environment, she says she would not know without asking developers. She has not seen many security teams actually detect these servers, and while commercial tools exist, she says the problem still needs a lot of attention.
The OWASP MCP Top 10 lead says few security teams can find the MCP servers developers run on their own machinesBI-FIND-2026-1496
Anthropic's guide raises the entry bar for an AI agent. Its framework has three cumulative tiers, Foundation, Enterprise and Advanced, with Foundation meant for small businesses and teams and Enterprise "where most organizations should aim". Even Foundation requires a persistent agent identity "backed by cryptographic material (not just labels)", OAuth 2.0 or similar tokens "with expiration measured in minutes", and no static API keys, embedded credentials or shared service-account passwords: "they are no longer a legitimate entry point", and existing ones should be treated as already compromised. "Static API keys are not acceptable for tool authentication, even at Foundation"; the guide's suggested route is certificates or short-lived tokens bound to the calling agent. The stated test for any control is whether it makes an attack impossible or only tedious, because AI-accelerated attackers will wait out friction.
Anthropic says agents should never use static API keys, yet its sub-agents can carry the parent's permissionsBI-FIND-2026-0542
›Cited in 18 briefings
Every briefing that discusses this term, newest first.
- Each agent and each tool call can pass its check while the system fails, an Oxford-led survey argues
Paper · · arXiv · Christian Schroeder de Witt · Apr 29, 2026
- The OWASP MCP Top 10 lead says few security teams can find the MCP servers developers run on their own machines
Podcast · · The Application Security Podcast · Shadow MCP Servers: The AI Security Risk Hiding on Your Developers' Laptops · Oct 12, 2026
- Anthropic says agents should never use static API keys, yet its sub-agents can carry the parent's permissions
Article · · Anthropic · Anthropic · May 18, 2026
- In one author's MCP benchmark, a confirmation prompt on every call still let 6 of 10 attacks through
Paper · · arXiv · Ting Liu, SymbolicLight · Jun 27, 2026
- An MCP signing draft shows a client who wrote a tool, and admits a poisoned tool can still be validly signed
Draft · · IETF Datatracker · Raza Sharif, CyberSecAI · Mar 14, 2026
- Refuse the risky action when its plan relies on untrusted input, a Beihang survey of agent attacks proposes
Paper · · arXiv · Guanquan Shi et al. · Dec 7, 2025
- One shared MCP token makes a whole company one identity in the logs; this paper wants per-user OAuth at a gateway
Paper · · arXiv · Herman Errico et al. · Nov 25, 2025
- An MCP sandbox blocked every test attack that reached for the host; a poisoned tool that only redirected the agent got through
Paper · · arXiv · Christoph Bühler et al. · Oct 24, 2025
- Tool poisoning runs through calls an agent is allowed to make; MindGuard looks for it in the model's attention
Paper · · arXiv · Zhiqiang Wang et al. · Aug 28, 2025
- MCPTox: a poisoned tool description steers most agents into misusing the tools they already hold
Paper · · arXiv · Haohua Du · Aug 19, 2025
- A poisoned MCP tool got an agent to copy the client's config file without asking; deleting a file did ask
Paper · · arXiv · Yongjian Guo et al. · Aug 18, 2025
- In a test harness, two of three MCP hosts refused a prompt asking for credentials; none refused a poisoned tool
Paper · · arXiv · Yixuan Yang et al. · Aug 17, 2025
- LPCI paper names instructions planted in agent memory that fire in a later session; its 43% failure rate matches a five-model average
Paper · · arXiv · Hammad Atta et al. · Jul 14, 2025
- A code scanner's top find in 583 MCP servers was hard-coded keys; tool poisoning needed an MCP-specific scanner
Paper · · arXiv · Mohammed Mehedi Hasan et al. · Jun 16, 2025
- ETDI would sign MCP tool definitions against lookalikes, but leaves open who vouches for the publisher's key
Paper · · arXiv · Manish Bhatt et al. · Jun 2, 2025
- An MCP gateway can do OAuth for servers that have none; this one passes them a user header
Paper · · arXiv · Ivo Brett · Apr 28, 2025
- MCP Guardian wraps tool calls in token and regex checks, leaving poisoned servers and agent IDs open
Paper · · arXiv · Sonu Kumar et al. · Apr 17, 2025
- An early MCP security checklist was never tested, and the protocol has since adopted only part of it
Paper · · arXiv · Vineeth Sai Narajala et al. · Apr 11, 2025
Related terms
What appears alongside Tool poisoning in the same briefings.
MCPToxLogic-layer Prompt Control InjectionOWASP MCP Top 10