Last updated
What is AI agent security testing?
AI agent security testing is adversarial testing of AI systems that take actions, not just produce text. An agent that can read your email, call your APIs, query your database or trigger payments has real-world reach, and every piece of untrusted content it processes is a potential instruction. Ironbark Cyber tests the full path: can an attacker influence what the agent sees, and can that influence be turned into an action that hurts you?
What do you test?
- Tool and function abuse. Every tool the agent can invoke, the privileges it runs with, and what an attacker can make it do through crafted input.
- Excessive agency. Actions the agent can take that it never should have been able to, the gap between "what we intended" and "what the permissions actually allow".
- Indirect prompt injection to action. Instructions planted in documents, tickets, emails or web content that the agent reads and obeys, chained through to a consequential tool call.
- Memory and context poisoning. Persistent manipulation of an agent's memory or knowledge base that changes its behaviour for later, legitimate users.
- Multi-agent trust boundaries. Confused-deputy attacks and privilege escalation through agent-to-agent calls in orchestrated systems.
- Human-in-the-loop bypass. Whether approval steps constrain the agent, or can be socially engineered or overwhelmed.
How do you test it?
We map the agent's trust graph, inputs, tools, permissions, memory and other agents, then attack each edge and chain the results. Testing is aligned to the OWASP LLM Top 10 (particularly excessive agency and insecure output handling) and MITRE ATLAS, and reported so the action impact of each finding is explicit, not left as "the model can be tricked".
What do we receive?
- An executive summary and an impact-ranked technical report with reproduction steps.
- A trust-graph diagram of the agent system with the attack paths marked.
- Remediation guidance covering permissions, tool design, input handling and human oversight.
- A debrief call, an attestation letter, and a free retest of fixed issues within 90 days.
How long does it take and what does it cost?
Typically 5–12 testing days depending on scope. Fixed-fee, indicatively from AU$20,000, quoted exactly within one business day of a scoping call.
FAQ
Frequently asked questions
How much does AI agent security testing cost?
Agent testing is fixed-fee, quoted within one business day of a scoping call. Cost tracks the number of tools the agent can invoke, how many agents interact, and whether the agent acts autonomously or with a human in the loop. Indicative pricing is published on this page.
What is an AI agent, for testing purposes?
An AI agent is an LLM given the ability to take actions, call APIs and tools, read and write data, spend money, or invoke other agents, often across multiple steps with memory. That autonomy is exactly what makes agents useful and what makes them dangerous: a prompt injection stops being a text problem and becomes an action problem.
What is the biggest risk with agentic systems?
Excessive agency combined with untrusted input. If an attacker can influence what the agent reads, and the agent can take consequential actions, the attacker can often chain the two: injected instructions in a document lead to a tool call that leaks data or moves money. We test that whole path, not just the prompt.
Do you test multi-agent systems?
Yes. Multi-agent systems add confused-deputy attacks, trust boundaries between agents, and privilege escalation through agent-to-agent calls. We map the trust graph and test each edge, because the weakest agent often becomes a pivot into the others.
How does this differ from an LLM penetration test?
LLM pentesting focuses on a model-backed application; agent security testing focuses on what the model can do, the tools, permissions, memory and orchestration around it. There is overlap, and for an agentic product we usually scope both together so injection findings are followed through to their real-world action impact.
How long does agent security testing take?
Typically 5–12 testing days depending on the number of tools and agents in scope, with the report delivered within a week of testing finishing. Critical findings are escalated live.
Contact
Talk to us
Tell us what you're trying to protect, secure or build. We'll come back with a plan.