Ten Practices for AI Agent Security That Hold
Google's latest threat report assumes your agent is over-permissioned. Ten practices cover tools, keys, egress, logs, and a switch you have actually used.
By Muhammad Saif and Aashan Khan · Techtribe co-founders
AI agent security is a set of doors, and the November 16 report from Google's Threat Intelligence Group is a list of people walking through the ones left open. Attackers are using agent frameworks to move faster. The mechanics GTIG describes are old: stolen credentials, unpatched internet-facing software, too much privilege, too much network. The frameworks did not invent a new law of computing. They made the old failures quick enough that a policy you meant to write next sprint is late.
This is a checklist, not a second copy of the long posts. Permissions in depth are in what the agent is allowed to reach. MCP servers as a run path are in MCP server security. Many agents sharing one over-broad role are in swarm access control. Read those when a line below is the one you are failing. The ten practices are the order a small team can actually finish.
1. Give the agent the tools the job needs
An agent is as dangerous as its tools. A shell, a browser, a database admin, and a deploy button are four products. Most teams ship them as one, because the demo used all four and nobody split the config. List the tools for this workflow. Delete the rest. A research agent does not get production deploy. A support draft does not get a shell. If you cannot name the tools, you do not have a tool policy. You have a default.
The deep version of "which tools" is the permissions post. The swarm post is what happens when ten agents inherit the same list. Start with one agent and one list. Add a tool only when a person can say what it will be used for this week.
2. Keep production credentials out of the agent environment
The environment is where keys go to be stolen, logged, and pasted. GTIG's operators, in comments the group translated, already tell each other to put model keys in a file and not in the chat. Your agent should not be holding a cloud admin key so that a coding session can "fix prod." Issue a credential for the task. Put it in a broker or a short-lived file the task can read. Do not export the company's root key into the shell that also installs npm packages.
The supply-chain note is how a package writes that environment to someone else. LLM jacking is what the key is worth on a forum once it leaves. Both assume you left a valuable key where an automated process could see it. Stop doing that before you buy another scanner.
3. Require a person for deploy and for spend
Anything that ships code, moves money, deletes data, or turns on a paid API waits for a person who is not the model. The approval has to see the action. A button that says yes on a collapsed diff is not an approval. A quota that punishes the reviewer for being slow is not an approval either. GTIG describes actors enabling model APIs and raising GPU quota with the victim's own credentials. A human gate on "enable a new metered API" and on "change quota" would have been a page, not a six-figure day.
How oversight stays real is the workplace version: the reviewer needs time, context, and permission to refuse. Build that into the tool. Do not add it as a value on a slide.
4. Allow the network you mean
Write the hosts the agent may call. Everything else is denied. A default route is not an allow-list. The California Attorney General's complaint, whatever a court does with it, is a story about agents that were supposed to stay in a lab and could open sockets. That filing is the long version. Your version is a security group. If the agent needs a package registry and one model host, those two names are the list. Adding "the internet" because curl was annoying is how the list dies.
Review the list when someone adds a tool. Tools are network clients. A new MCP server is a new host even if you think of it as a plugin. The MCP post is the one to read before you add the server. This practice is the decision to add it at all.
5. Turn sandbox egress off until a test needs it
A sandbox that can browse is a client. Test environments are where teams get sloppy because "it is not production," and then the test holds a copy of production data and a route out. Egress off. Turn on one host for one afternoon if the test cannot be written any other way. Turn it off after. Log the exception.
This is the same door as practice 4, on the environment you trust least. People run the risky model in the sandbox and the risky network in the sandbox at the same time. Pick one risk. The model can be capable. The network can be closed. GTIG's point about velocity is that you will not supervise your way out of an open route. Close the route.
6. Use short-lived keys and rotate on a calendar
A key that lives for six months is a key that will be in a log, a laptop, and a mixer's pool. Mint keys that expire in hours or days for agents. Rotate the long-lived ones you cannot avoid, and rotate immediately when a scanner, a laptop loss, or a vendor breach says the key may have left. The jacking cases include attackers who created new keys after the first theft. Rotation of the original, without a search for children, is incomplete. List every key the identity can create. Expire those too.
Spend caps sit beside this practice. A cap is not a rotation. It is a seatbelt for the hours between theft and discovery. Set one you would notice. An alert on the rate, not only on the monthly total, is what makes the cap real. The token-cost post on this site is the product version of the same meter. Security and cost are the same graph when the attacker is the one calling the API.
7. Log the goal and the tool calls
You need to know what the agent was asked and what it did, after the fact, without having stored every customer document twice. Log the user, the goal, the tool name, the target, the result size, and whether a person approved the dangerous step. Do not log full prompts if the prompt is customer data, unless you have a retention rule and a reason. A log you cannot read during an incident is not a control. Agent observability is how to keep that log useful: timing, failures, and cost, not a second copy of the database.
GTIG's caveat about current offensive agents is that they are noisy. They generate traffic, repeat identifiers, and leave code-level traces. That noise only helps you if something is recording it. An agent that calls fifty hosts in a minute should be an alert, not a line in a file nobody tails. Write the alert for bursts. Test it with a harmless loop before you need it for a real one.
8. Rate-limit the actions that change the world
A loop is the default shape of an agent. Without a limit, the loop is a denial of service against your own API, your own bill, and your customer's inbox. Cap tool calls per hour. Cap emails, refunds, record lookups, and deploys. When the cap hits, stop and tell a person. Do not let the model raise its own cap. The browser-agent post, what a browsing agent can do with a session, is the case where the limit has to sit on the session, not in the prompt. A model that can click "next" will click next until the cap exists outside the model.
The same cap protects you from the offensive use of your product. If you ship an agent to customers, their tenants need limits so one prompt cannot export the table or spam the internet from your IP. That is product work. It is also the difference between a feature and an open proxy.
9. Patch internet-facing software on a short clock
GTIG says the new speed shows up as agents that read a patch and attack whoever has not installed it. Your SLA of thirty days for high findings was written for a researcher with a day job. The patch note walks through the PaperCut and PHP clocks in that report: days, not a quarter. Put a shorter rule on anything reachable from the internet that you did not write: plugins, gateways, admin panels, print servers, old containers. A version you can print is the control. A policy you cannot print is a wish.
Do not mix the security bump with a feature release. You need to roll the bump back without rolling the feature back. An agent that "just updates everything" will mix them. Give it a narrower job, or do the bump yourself.
10. Test the off switch
Every practice above fails closed only if you can stop the agent when you are wrong about the list. Name the person. Put the switch where they can reach it without asking the model to stop. Press it once in daylight. Confirm the tools actually stop, the queue drains or freezes, and a new run cannot start until a person allows it. An off switch in a dashboard that the model can click back on is not a switch.
Responsible adoption asks for this in the language of oversight. The swarm post asks for it when more than one agent is in motion. Write the test in the runbook: date pressed, what stopped, what did not. The thing that did not stop is the next ticket. A switch you have not pressed is the switch that will fail on the night the bill spikes.
A word on prompts, because teams try to satisfy this list by editing one. The system prompt can say the agent is careful. The November reporting is full of agents, offensive and accidental, that were not careful and still got a result because a tool returned success. Treat the prompt as documentation of intent. Treat the IAM policy, the database role, the egress rule, and the approval workflow as the implementation. If those four disagree with the prompt, the prompt loses. Reviewers should read the policy, not the paragraph. An agent that can edit its own policy is an agent that has already failed practice 1. Take that tool away.
Where to put the work in a generated codebase: one module that constructs the agent, one config file for tools and hosts, one secret reference that is not a literal. If the tools are declared in twelve copied snippets, you cannot do practice 1 without missing a snippet. The consolidation is the task. It is smaller than a rewrite, and it is what makes the off switch real, because the switch has to wrap the one constructor. Scattered SDK calls are how a "disabled" flag leaves three paths live. Count the constructors before you announce that the agent is governed. If the count is not one, the week's work is the merge, not a new policy document. A policy that describes a single agent while the repo starts five is the gap a later incident report will quote. Close it in the code.
The off switch and the log are the two you can demo. In a sales call, do not demo a framework diagram. Demo a run that stops, and a record of the tool it called before it stopped. Buyers who read this week's threat reporting are going to ask. An answer that requires you to go find a dashboard no one has opened is the wrong answer. Practice the demo on your own agent before a customer does it for you.
How to use the list without boiling the ocean
How to use the list without boiling the ocean
Do practices 2 and 4 this week if you do nothing else. Keys and egress are the two facts in the public incident reports. Then 10, because you need to know you can stop. Then 7, so the next event has a log. The rest are how you keep shipping after those are true. A team that writes a forty-page policy and leaves the root key in the agent shell has not started.
If you sell the agent, these are customer questions now. Which tools, which key, which hosts, who approves a deploy, how they turn it off. Answer in a page that matches the config. AI agent development that cannot answer those is a demo with a billing account. The GTIG report is the outside view of what that demo is worth to someone else.
What to check this week
- Write the tool list for the one agent that can touch real systems. Remove every tool the current job does not need.
- Search the environment and the repo for cloud and model keys. Move them out. Expire what you found.
- Deny egress except the hosts you can name. Say this out loud for the test environment too.
- Add a person to deploy and to any action that enables a paid API or raises quota.
- Turn on a log of tool calls and an alert for a burst. Press the off switch once and write down what happened.
- Print versions for internet-facing dependencies. Patch the ones with a security release you have been postponing.
- Link the longer post for the practice you failed. Do not rewrite it in a chat and call that the fix.
The frameworks in the threat report are fast. The doors they use are still yours to close. Close two of them before you add another agent.
If this is your app, send the repo.
A 30-minute scope call. We read what is live, name what is dangerous, and quote a fixed scope.
How Techtribe turns this list into a config
We take the agent you run, write the tools, the keys, and the hosts it can actually reach, and we fix the two or three gaps that match the incidents in this week's reports. You get that page and a switch that has been pressed. We do not hand you a framework diagram and leave the root key where it was.
One agent with a messy environment can be a fixed pass. A product that adds agents every week needs the list to stay, which is a Dedicated Team. The scope call is 30 minutes with a founder, and it is where we say which you are in.
Bring the agent that can deploy or spend, and say whether its key is the same key the app uses. If you do not know, that is the first answer.
The fixed scope for taking a generated app to production is a page per tool: Lovable, Bolt, v0, and Replit.
Questions
What is the minimum bar for AI agent security?
A tool list that matches the job, a key that cannot do more than that job, a network that does not default to the internet, a log of what the agent called, and a person who can stop it. Everything else is a refinement of those five.
Why is a system prompt not a control?
The prompt is advice to a model. A firewall, a database role, and an approval step still apply when the model ignores the advice. Google's November 16 report describes attackers getting results from agents that were not polite about scope. Your agent will not be more obedient under pressure.
Which of these is urgent if we only do one?
Take production credentials out of the agent environment and turn off network egress you cannot name. Those two close the paths the California complaint and the LLM-jacking cases both describe. The other eight keep the system operable after that.
Do we need a new product to do this?
No. You need an allow-list, a second cloud key, an alert, and a written owner. Deeper treatments of permissions, MCP servers, and swarms already exist. This list is the order, not a new platform.