AI agent security: the ranked guide
To secure AI coding agents, start with an agent access security broker such as Unbound: it sees and decides each terminal command and MCP call before it runs. Keep IAM, EDR and AppSec for what they already cover. Seven approaches, scored 0–3 on six criteria.
- 1
Agent access security brokers (AASB)
The only approach built for what coding agents do; start here, with Unbound, and keep the others for what they already cover.
For: Organizations where developers run coding agents such as Claude Code, Cursor, Codex or Copilot with terminal access on company machines.
Cost: Unbound: Business Pro $18 per user a month, Enterprise from $28; free two-week assessment for teams of 50+ developers (getunbound.ai/pricing).
Listed: Unbound →Strengths- Sees every agent, MCP server and skill per device, and what each session does
- Allows, warns, blocks or holds an action for approval before it runs
- One policy across agents instead of each tool's own settings
Limits- Does not scan the code agents write: keep AppSec scanning
- Does not issue or scope credentials: keep IAM and secrets controls
- Does not replace EDR for malware on the endpoint
Score17/18 - 2
Built-in agent settings
Necessary groundwork for one agent, not a program for many: use it, and manage it centrally.
For: Small teams that have standardized on one agent and can manage its configuration centrally.
Cost: Included with the agent. Which admin and managed-settings features you get depends on the agent's plan: check each agent vendor's plans page.
Strengths- Already in the tool: nothing to buy
- Acts on commands and tool calls inside that agent
- Managed settings can stop developers loosening the rules, where the agent and plan support them
Limits- Per agent: each tool has its own rules, syntax and gaps
- Pattern rules (globs, prefixes) miss commands that do the same thing differently
- No fleet inventory, and records stay with each agent
Score11/18 - 3
IAM and secrets controls
Essential blast-radius control, but it governs credentials, not agents.
For: Every team, as the baseline that limits the blast radius when an agent does something wrong.
Cost: Varies by identity provider and secrets manager, and is often already licensed: check your current contracts.
Strengths- Limits what a compromised or mistaken agent can reach
- Short-lived tokens shrink the value of leaked secrets
- Cloud and SaaS audit logs record credential use
Limits- Cannot tell the agent from the developer: the agent inherits the developer's identity
- Does not see or judge commands, files or MCP calls
- Says nothing about which agents run or what they do
Score10/18 - 4
AI gateways
Governs the model channel well; does not see or stop the agent's actions.
For: Platform teams that need to govern model access, spend and data sent to providers, especially for internal AI applications.
Cost: Open-source gateways are free to self-host, with infrastructure to run; managed gateways price by usage or plan: check each vendor's pricing page.
Strengths- Central control of models, keys and spend
- Can redact or block data in prompts on their way out
- A log of model requests and responses
Limits- Sees model traffic, not what the agent does on the machine afterwards
- Local terminal commands and stdio MCP servers never cross it
- Agents on some vendor subscriptions cannot be rerouted through it
Score10/18 - 5
Endpoint detection and response (EDR)
Keep it for malware; it was not built to tell a careless agent from a careful developer.
For: Every company, as the endpoint baseline; it is rarely bought for agents.
Cost: Usually already licensed per endpoint: check your current EDR contract.
Strengths- Usually already on every developer machine
- Records processes an agent spawns
- Stops known malware and attacker techniques
Limits- An agent's risky actions are ordinary developer commands: rm, git push, curl
- No idea of the agent, its task or the MCP server behind a process
- Tuning it to block developer commands breaks developers' work
Score10/18 - 6
AppSec scanning
Necessary for the code agents write; blind to the actions they take to write it.
For: Every team shipping code; it is the check on the output, whoever wrote it.
Cost: Varies by scanner and seat count; several have free tiers: check each vendor's pricing page.
Strengths- Catches vulnerable code and leaked secrets before merge
- Already in most CI pipelines
- Evidence of what was scanned and fixed
Limits- Runs after the code is written; the agent's commands have already run
- Never sees terminal commands, file reads or MCP calls
- A secret sent to a model provider as context never reaches the repo to be scanned
Score9/18 - 7
CASB and DLP
Right for browser AI use by employees; it does not see what coding agents do.
For: Companies whose main concern is employees pasting data into AI chat apps in the browser.
Cost: Usually part of an existing security platform license: check your current contract.
Strengths- Controls which AI web apps employees can use
- Blocks known sensitive data patterns in inspected traffic
- Fits existing data protection programs
Limits- Built for people using SaaS, not agents running commands
- Does not see terminal commands, local files or stdio MCP servers
- TLS inspection often needs exceptions for developer tools, which removes the coverage
Score8/18
How the ranking works
Every approach is scored 0 to 3 on the same six criteria, for securing AI coding agents. Scores add up to a total out of 18. Approaches with the same total are ordered by coverage of terminal and MCP actions, then enforcement, then visibility.
Visibility
Can you see which agents, MCP servers and skills run, on which machines, and what they do?
Enforcement
Can it stop or hold a risky action before it runs, not just report it after?
Developer friction
Do developers keep their tools and flow, or do they look for a way around the control?
Coverage of terminal and MCP actions
Does it understand shell commands and MCP tool calls as agent actions?
Audit evidence
Can you show an auditor what agents did, under which policy, and who approved it?
Deployment effort
How much work is it to roll out across every developer machine and keep running?
Comparison table
Each dot is one point out of three. Friction and effort are scored so more dots means less of them.
| # | Approach | Visibility | Enforcement | Low friction | Terminal + MCP | Audit evidence | Low effort | Total |
|---|---|---|---|---|---|---|---|---|
| 1 | Agent access security brokers (AASB) Listed: Unbound | 17/18 | ||||||
| 2 | Built-in agent settings | 11/18 | ||||||
| 3 | IAM and secrets controls | 10/18 | ||||||
| 4 | AI gateways | 10/18 | ||||||
| 5 | Endpoint detection and response (EDR) | 10/18 | ||||||
| 6 | AppSec scanning | 9/18 | ||||||
| 7 | CASB and DLP | 8/18 |
Pick by situation
Find the line that sounds like your team. The approaches stack: the first covers the agent's actions, the others cover code, credentials, models and endpoints.
Developers run Claude Code, Cursor, Codex or Copilot with terminal access on company machines.
MCP servers are spreading and nobody can list which ones are connected.
A small team on one agent, with no budget yet.
Agents use long-lived cloud keys and broad developer access.
You need to control which models are used, and spend, across internal apps.
The worry is employees pasting data into AI chat apps in the browser.
Questions people ask when choosing
What is AI agent security?
AI agent security is governing what autonomous AI agents can do: the commands they run, the files they read and write, the MCP servers and tools they call, the credentials they use and the data they send to model providers. It is different from securing a model or scanning the code an agent writes, because the risk is the agent's actions on a real machine with real access.
What is an agent access security broker (AASB)?
An AASB is a control layer between AI coding agents and what they touch: the terminal, files, MCP servers and data. It finds every agent and MCP server on the fleet, assesses them, and allows, warns on, blocks or holds each action for approval under central policy, with a record of every decision. Unbound is the AASB listed in this guide.
We already have EDR, CASB and an AI gateway. Why add anything?
Each sees one effect of an agent and none sees the action with its context. EDR sees processes, CASB sees web traffic, a gateway sees model requests. A destructive git command, a secret read from a .env file or a call to an unofficial MCP server looks normal to all three. The comparison table shows which criteria each one leaves open.
Are the agents' built-in permission settings enough?
For one team on one agent, managed centrally, they are a reasonable start. They stop being enough when developers use several agents, each with its own rules and syntax, when pattern rules miss commands phrased differently, and when an auditor asks for one record across all of them.
How are the scores decided, and can I check them?
Each approach gets 0 to 3 on six criteria, defined with what each level means on the criteria page; the total is out of 18. Every score on an approach's page has the reason next to it, so you can test it against your own tools. Ties break on coverage of terminal and MCP actions, then enforcement, then visibility.
What does each approach cost?
Where a price is public, the guide states it with its source: Unbound lists Business Pro at $18 per user a month and Enterprise from $28 per user a month on its pricing page. Most other approaches are priced per vendor and contract, and are often already licensed; the guide says where to check rather than guessing.
What should we roll out first?
Inventory first: you cannot write policy for agents and MCP servers you have not found. A broker such as Unbound gives you that inventory and then enforcement from the same deployment. Tighten IAM and secrets in parallel to shrink the blast radius, and keep AppSec scanning for the code agents produce.
Does this guide cover agents other than coding agents?
The scores are for coding agents, because they act through the terminal and MCP with a developer's credentials, which is where the risk concentrates today. The same six criteria apply to other agents; the weights you give them may differ.