Threat Models for Tool-Using Agents
- Publication date
- Oct 16, 2026
- Author
- SkyNet AI Research
- Status
- Active Research
Abstract
An agent that can call tools inherits the attack surface of every tool it can reach. This paper catalogues threat classes specific to tool-using systems, including instruction injection through retrieved content, privilege escalation across tool boundaries, and exfiltration through legitimate channels.
Technical documentation
The inherited surface
When an agent can read email, query a database, and write to a ticket system, a compromised instruction can chain those capabilities. The risk is not the model's intent but the composition of permissions.
Threat classes
We describe four classes: injection through untrusted retrieved content, confused-deputy behavior where the agent acts with more authority than the requester, cross-boundary data movement, and action repetition under adversarial input.
Controls that matter
Least-privilege tool scopes, explicit approval for irreversible actions, content provenance on retrieved material, and an immutable execution log. None of these are model improvements; all are architecture.
Residual risk
Controls reduce but do not eliminate risk. Deployments should state which residual risks are accepted and why.
References
- [1]Greshake et al., Not What You've Signed Up For, 2023
- [2]OWASP Top 10 for LLM Applications, 2025