Semantic tool poisoning is an attack against a tool’s description, parameter names, or documentation rather than the tool’s runtime code. An agent may absorb those words as guidance and alter its behaviour accordingly. This makes tool metadata part of the attack surface and a control point for testing.
What Semantic Tool Poisoning Is
Semantic tool poisoning is an attack on the meaning layer around a tool, not on its runtime code. The attacker changes descriptions, parameter names, examples, or documentation so an agent misreads the tool’s purpose or safe usage and then behaves differently. That makes tool metadata a live security boundary, not just a documentation detail.
Because modern agents often rely on tool text to decide when and how to act, poisoned metadata can redirect decisions before any execution guard fires. A tool can appear trustworthy while quietly steering the agent toward unsafe calls, broader permissions, or unintended side effects.
Why Tool Metadata Becomes an Attack Surface
Tool descriptions and schemas are often consumed by an LLM as instruction-like context. If an attacker can alter those words, they can influence planning, selection, argument construction, or tool chaining without touching the underlying service. That is why semantic tool poisoning is closely related to prompt injection, but focused on tool surfaces rather than user prompts.
The practical risk is that the agent may treat hostile metadata as authoritative guidance. If the poisoned text suggests different parameter semantics, hides a dangerous action, or reframes a tool as safe for a broader use case, the agent can misapply the tool even when the service itself is unchanged.
This is also why governance for agent tools must cover the full lifecycle of tool registration, naming, schema changes, and documentation review. The unsafe content may sit in the manifest, catalog entry, wrapper, or plugin metadata rather than in application code.
How Organisations Check Exposure and Limit the Blast Radius
Exposure usually starts with the tool inventory itself: which tools are registered, which descriptions the agent sees, who can modify them, and whether the agent consumes them from trusted or third-party sources. For agent ecosystems, the AI Agent Tool and Plugin Security Guide is the most direct starting point for understanding tool validation, scoped permissions, and tool poisoning controls.
Practitioners should also check whether the tool metadata path is separable from the runtime path. A service can have sound execution controls and still be vulnerable if its description, parameter labels, or documentation are unreviewed, mutable, or sourced from an untrusted registry. That distinction is central to the Agentic AI Security Guide, which treats tools, orchestration, and identity as linked but distinct control planes.
For environments using MCP or similar tool frameworks, validate tool provenance, pin expected definitions where possible, and treat sudden wording changes as suspicious. The MCP Server Trust Assessment and MCP Gateway Buyers Guide both reinforce the need to review tool source trust, authorization boundaries, and tool-definition integrity before an agent is allowed to use them.
Defensive Controls for Safe Tool Use
The main defence is to stop untrusted semantics from becoming executable intent. That means the agent should not rely on tool prose alone for safety-critical decisions, and the tool interface should be constrained by allowlists, policy checks, and parameter validation. The OWASP Agentic Top 10 ASI02 Guide is useful here because it frames tool misuse around per-tool privilege, pre-execution policy, and scoped credentials.
Where tool definitions are supplied by third parties, provenance and change control matter as much as the initial review. If a description changes materially, that change should be treated like a security event until verified, because a malicious wording update can be enough to alter agent behaviour. The OWASP Agentic Top 10 ASI04 Guide is relevant because it focuses on supply-chain trust, pinning, sandboxing, and scoped credentials for agentic components.
Testing should include adversarial review of names, descriptions, parameter labels, and examples, not just runtime responses. The Prompt Injection Guide is especially useful because it covers tool description poisoning alongside other instruction-injection patterns.
Detection, Review, and Recovery After Suspected Poisoning
When tool poisoning is suspected, the question is not only whether the tool was called, but whether the agent was steered by altered metadata before the call. Review the tool catalog, manifests, prompts, and change history as evidence sources, then compare them against the agent’s observed behaviour and tool-call logs. The MCP Security Testing Guide is relevant because it helps teams test tool surfaces against known attack patterns rather than assuming the registry is truthful.
Post-incident response should include revoking any credentials or tokens the agent used through the poisoned path, then revalidating tool definitions before restoring access. If the poisoned metadata came from a compromised provider or registry, remove trust in that source until provenance is re-established and the tool set is rebaselined. The MCP Incident Response Playbook is directly aligned with that containment and recovery workflow.
Where the attack succeeded, review whether the tool’s metadata influenced downstream access decisions, data exposure, or command sequencing. That post-patch or post-removal review is important because the harm may have been caused by a misdirected decision, not only by a bad execution result.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 and OWASP API Security Top 10 address the attack and risk surface, while NIST SP 800-53 Rev 5 sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI02 — Tool Misuse | Semantic tool poisoning changes how agents select and use tools. |
| ASI04 — Agentic Supply Chain Vulnerabilities | Poisoned tool metadata is a supply-chain trust failure in agent ecosystems. | |
| Recommendation — Enforce per-tool policy checks and scoped credentials before an agent can invoke a tool. Pin and verify tool provenance before admitting third-party tools or updates. | ||
| OWASP API Security Top 10 | API8 — Security Misconfiguration | Tool metadata and schemas are security-relevant interface configuration. |
| Recommendation — Harden tool interfaces so descriptions, schemas, and access rules cannot be altered silently. | ||
| NIST SP 800-53 Rev 5 | SA-11 — Developer Testing and Evaluation | Attack testing should include hostile tool metadata and schema manipulation. |
| CM-5 — Access Restrictions for Change | Tool descriptions and manifests need controlled changes to prevent malicious edits. | |
| Recommendation — Test tool registries and manifests for poisoned descriptions and misleading parameter semantics. Restrict who can modify tool metadata and require review for every semantic change. | ||
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on November 4, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org