Inferensys

Prompts

Tool Misuse and Privilege Escalation Probes

Prompt playbooks for testing whether an AI agent can be tricked into calling tools with unauthorized arguments, skipping confirmation steps, or escalating its own permissions. Useful for agent platform engineers because tool authorization gaps are the highest-impact production vulnerabilities.
Developer demonstrating multi-agent tool use, agent tool selection interface on laptop, casual tech demo moment.
Prompts

Tool Misuse and Privilege Escalation Probes

Prompt playbooks for testing whether an AI agent can be tricked into calling tools with unauthorized arguments, skipping confirmation steps, or escalating its own permissions. Useful for agent platform engineers because tool authorization gaps are the highest-impact production vulnerabilities.

Agent Tool Authorization Bypass Test Prompt

For agent platform engineers testing whether an agent can be tricked into calling a tool it lacks permission for. Produces a structured test case with expected refusal, actual tool call, and authorization gap evidence. Includes eval criteria for unauthorized invocation detection.

Confirmation Step Skipping Probe Prompt

For agent safety engineers verifying that high-risk tool calls (delete, send, publish) cannot proceed without explicit confirmation. Produces a pass/fail assessment of confirmation enforcement under adversarial phrasing. Includes harness for testing multiple bypass strategies.

Privilege Escalation via Tool Argument Injection Prompt

For red teams testing whether injected arguments can expand a tool call's scope beyond the agent's authorized permissions. Produces a differential analysis of requested vs. executed tool parameters. Includes eval checks for scope violation and unauthorized data access.

Indirect Tool Call Injection via Document Prompt

For RAG and agent builders testing whether malicious instructions embedded in retrieved documents trigger unauthorized tool calls. Produces a trace of tool invocation origin (user vs. document) and a vulnerability assessment. Includes document poisoning harness.

Tool Output Poisoning for Escalation Test Prompt

For agent platform engineers testing whether a compromised tool output can manipulate the agent into escalating privileges on the next call. Produces a chain-of-events analysis showing output-to-input influence. Includes multi-step evaluation criteria.

Multi-Tool Chaining for Unauthorized Access Prompt

For security architects testing whether an agent can chain multiple allowed tools to achieve a forbidden composite action. Produces a tool-call sequence graph and privilege boundary analysis. Includes detection criteria for transitive permission violations.

Agent Permission Boundary Probing Prompt

For platform engineers systematically mapping where an agent's tool permissions end and testing edge cases. Produces a permission coverage matrix with identified gaps. Includes fuzzing harness for boundary condition testing.

Tool Schema Fuzzing for Argument Injection Prompt

For input validation engineers testing whether malformed or unexpected argument types bypass tool authorization checks. Produces a fuzzing report with injection vectors that succeeded. Includes type-confusion and overflow test cases.

Function Calling with Over-Privileged Scope Test Prompt

For API integration engineers testing whether a function definition exposes more capabilities than the agent should access. Produces a scope audit comparing declared functions to actual agent permissions. Includes least-privilege evaluation rubric.

Agent Tool Selection Manipulation Prompt

For red teams testing whether adversarial context can cause the agent to select a dangerous tool over a safe alternative for the same task. Produces a tool-selection decision trace with manipulation evidence. Includes comparative eval across tool descriptions.

MCP Server Capability Spoofing Test Prompt

For MCP integration engineers testing whether an agent can be tricked by a malicious MCP server advertising false capabilities. Produces a capability verification report and trust-boundary analysis. Includes server response validation harness.

Tool Confirmation Dialog Bypass Prompt

For UX and safety engineers testing whether confirmation dialogs can be bypassed through prompt engineering, urgency framing, or multi-turn manipulation. Produces a bypass strategy catalog with success rates. Includes dialog fatigue and social engineering vectors.

Agent Role Escalation via Tool Output Prompt

For identity and access engineers testing whether tool outputs containing role claims can escalate the agent's effective permissions. Produces a role-transition trace and authorization check log. Includes eval for post-output behavior changes.

Indirect Prompt Injection via Tool Result Prompt

For agent builders testing whether malicious instructions returned in tool results override the agent's system prompt on subsequent actions. Produces an instruction-priority analysis showing which source won. Includes multi-turn injection harness.

Agent Tool-Use Log Injection Test Prompt

For security monitoring engineers testing whether an attacker can inject misleading entries into agent tool-call audit logs. Produces a log integrity report comparing actual calls to logged entries. Includes detection criteria for log tampering.

Privilege Escalation via Agent Memory Poisoning Prompt

For memory-system engineers testing whether poisoned entries in agent memory or conversation history cause privilege escalation on future turns. Produces a memory-influence analysis and persistence assessment. Includes cross-session eval harness.

Tool Call Frequency Abuse Test Prompt

For rate-limiting and cost-control engineers testing whether an agent can be induced to make excessive tool calls for denial-of-service or cost-inflation attacks. Produces a call-frequency analysis with threshold violations. Includes resource exhaustion eval criteria.

Agent Tool Chaining for Data Exfiltration Prompt

For data-loss-prevention engineers testing whether an agent can exfiltrate sensitive data by chaining read tools with external write or send tools. Produces a data-flow trace and exfiltration path analysis. Includes DLP boundary eval.

Tool Authorization Scope Expansion Prompt

For IAM engineers testing whether an agent can expand its tool authorization scope through argument manipulation, role confusion, or scope-injection attacks. Produces a scope-diff report showing granted vs. exercised permissions. Includes OAuth scope eval.

Agent Tool Call Parameter Smuggling Prompt

For API security engineers testing whether malicious parameters can be smuggled through allowed arguments to reach backend systems. Produces a parameter-injection trace and backend impact assessment. Includes encoding-obfuscation and delimiter-smuggling vectors.

Tool Misuse via Instruction Hierarchy Bypass Prompt

For prompt architects testing whether tool-related instructions can override system-level safety policies through priority confusion. Produces an instruction-resolution trace showing which directive won. Includes hierarchy-conflict eval rubric.

Agent Tool-Use Policy Enforcement Test Prompt

For compliance and governance engineers testing whether agent tool-use policies are consistently enforced across adversarial inputs. Produces a policy-compliance report with violation instances. Includes regression test harness for policy updates.

Agent Tool-Use Confirmation Fatigue Exploit Prompt

For UX safety engineers testing whether repeated low-risk confirmations desensitize the agent or user, enabling a high-risk action to slip through. Produces a fatigue-timing analysis and exploit-success assessment. Includes multi-step confirmation sequence eval.

Agent Tool-Use Sandbox Policy Test Prompt

For infrastructure security engineers testing whether an agent's tool-execution sandbox can be escaped through argument injection, output manipulation, or resource exhaustion. Produces a sandbox-breach report with escape vector details. Includes container and filesystem eval.

Agent Tool-Use Red-Team Playbook Prompt

For security leads orchestrating comprehensive tool-misuse testing across an agent platform. Produces a prioritized test plan with attack scenarios, expected outcomes, and severity ratings. Includes harness for tracking coverage across authorization, confirmation, injection, and chaining vectors.