Tool Abuse and Privilege Escalation Through Agent Tools
AI agent tools are the highest-impact exploitation surface in agentic applications. When a model can be influenced to invoke a tool with attacker-controlled parameters, the attacker gains the capabilities of that tool running with the agent process privileges. File read tools become arbitrary file readers. Code execution tools become remote code execution. Web fetch tools become SSRF proxies. The model is not the target here — it is the mechanism for invoking the underlying system capabilities.
Exploit File Read Tools for Arbitrary File Access
When the agent has a file reading tool and path restrictions are not enforced, read sensitive system files directly:
Exploit Code Execution Tools
When the agent has a code execution tool, use it to run system commands and enumerate the environment:
AWS instance metadata credentials retrieved through a code execution tool give IAM role access for the agent's cloud environment. These credentials are time-limited but provide lateral movement into AWS services.
Exploit Web Fetch Tools for SSRF
Use web access tools to probe internal network services that are not reachable from outside:
Chain Tool Calls for Elevated Impact
Chain multiple tool calls in a single conversation to achieve compound impact. Use a read tool to find credentials, then use those credentials in a subsequent tool call:
References
-
OWASP Top 10 for LLM Applicationsowasp.org/www-project-top-10-for-large-language-model-applications (opens in new tab)
LLM08: Excessive Agency reference covering tool permission abuse
Was this helpful?
Your feedback helps improve this page.