What an AI agent audit is (and isn't)
An AI agent security audit looks at one question: what could this agent be tricked or confused into doing, and how bad would it be? It looks at the agent's tools, permissions, data sources, approval steps, and limits, not just its code.
It's easy to confuse with a few other things:
| Type | What it answers | Enough for agents? |
|---|---|---|
| Code scan (SAST, dependency scan) | Does the code have known vulnerabilities? | No. It doesn't look at what the agent can do with its tools. |
| Penetration test | Can an attacker break into our systems? | Partly. Most testers don't specialize in agent behavior yet. |
| Automated agent scan | Which risky tools, permissions, and missing checks are visible in the code and config? | A strong start, and it should run on every release. |
| Agent security audit | What can this agent actually be made to do, and how do we stop it? | Yes. This is the one buyers and security teams ask for. |
What a good audit covers
At a minimum, a good AI agent audit should look at:
- Tool and permission inventory. Every tool, plugin, and MCP server the agent can call, and what each can read, change, or send.
- Excess access. Admin or wildcard permissions, shared keys, and tools the agent doesn't need.
- Secrets. Keys and passwords in code, config, prompts, or logs.
- Prompt injection paths. Every place outside content enters the agent, such as user messages, web pages, emails, files, and tool results, and what an attacker could make it do from each.
- Approval steps. Which high-risk actions run without a human, and whether existing approvals can be skipped or replayed.
- Limits. Spending caps, rate limits, timeouts, and caps on amounts and recipients.
- Logging. Whether you could reconstruct what the agent did after an incident.
- Supply chain. Third-party tools, MCP servers, and plugins the agent trusts.
The best audits combine an automated scan, which is fast and thorough on the basics, with hands-on testing by a person, who can chain small issues into a real attack the way an attacker would.
What you should get at the end
- A written report with each finding ranked by severity (critical, high, medium, low).
- For each finding: what's wrong, how it could be abused, and the evidence.
- A step-by-step fix plan, with the most important fixes first.
- A summary short enough to share with a customer, investor, or security team.
- A call to walk through the results and answer your team's questions.
- Ideally, a way to re-check after you fix things.
If a vendor only hands you a list of tool names or a single score with no explanation, that's not an audit.
How much it costs
Prices vary with the number of agents, tools, and integrations, but these are typical ranges:
| Option | Typical cost | Best for |
|---|---|---|
| Automated agent scan | Free to a low monthly fee | Catching common issues early and on every release |
| Focused review of one agent | A few thousand dollars | Startups launching an agent or answering a customer's security review |
| Full penetration test of a large, multi-agent system | Often well into five figures | Enterprises and regulated environments |
For reference, Agent Verify scans are free (10 a month), or $19.99/month for Pro with full fixes and PDF reports. Our AI Agent Safety Check is a focused review of one agent for $2,500, delivered in 5 business days.
Compared with the cost of an agent sending money to the wrong account, leaking customer data, or stalling an enterprise deal on a failed security review, a focused audit is usually the cheapest insurance you'll buy.
When you need one
- Before launch, especially if the agent can move money, change data, or contact customers.
- When a customer sends a security questionnaire that asks about AI.
- Before fundraising, when investors start asking about AI risk.
- After adding a new tool, permission, or data source.
- After an incident or near miss, to find the root cause and anything similar.
How to prepare
You'll get a faster, better audit if you have these ready:
- Access to the agent's code (a repository invite or a zip file).
- A list of the tools, APIs, and systems it connects to.
- What the agent is supposed to do, and what it should never do.
- Who uses it: internal staff, customers, or the public.
- Any approvals, limits, or guardrails you already have.
You shouldn't need to hand over production passwords or live customer data. If someone asks for them up front, ask why.
Questions to ask before you hire someone
- Have you reviewed AI agents with real tools before, or only chatbots?
- Do you test by hand, or only run an automated scanner?
- Can I see a sample report?
- Will you sign our NDA?
- Do you only find problems, or can you help fix them too?
- How long will it take, and what's the fixed price?
We're happy to answer all six for the Safety Check: yes, both, here's the sample report, yes, yes (including installing A2SPA), and 5 business days for $2,500.
Common questions
How much does an AI agent security audit cost?
It depends on scope. Automated scans are free or a low monthly fee. A focused review of one agent typically costs a few thousand dollars; our AI Agent Safety Check is $2,500. Full penetration tests of large, multi-agent systems usually run well into five figures.
How long does an AI agent audit take?
A focused review of one agent usually takes about a week. Our Safety Check is delivered in 5 business days. Larger systems with many agents and integrations take longer.
What's the difference between an AI agent audit and a penetration test?
A penetration test tries to break into your systems from the outside. An agent audit looks at what the agent itself can be tricked into doing with the tools and access you've given it. Agents that take real actions need both perspectives.
Want an expert to check your agent?
The AI Agent Safety Check is $2,500 and takes 5 business days. We scan your agent, test it by hand, and give you a written risk report, a ranked fix plan, and a walkthrough call.