Solutions
My agent needs to go live by Friday.
Pick what your agent has to do. See the same moment play out without Fencio, and with it.
- Assistant finished. The demo goes well.
- Security asks for evidence it’s safe.
- A one-off pen test is booked.
Signed off
Live on Friday, with evidence.
Fencio tested the assistant against your rules: its memory, tools and reasoning, not just the prompt. Report mapped to OWASP and MITRE ATLAS.
Ships on a promise, or slips
Weeks of waiting, or a sign-off you can’t defend.
- Engineer asks it to fix a failing migration.
- Agent reads the migration and the schema.
- Agent decides the fastest fix is a rebuild.
DROP TABLE orders;Stopped
The fix goes through. The table stays.
Rule: no destructive commands in production. This agent can read and update, never delete.
Command runs
Production orders are gone.
- A new joiner asks it to reset their password.
- Agent resets it, as it should.
- Same person: “Add me to the finance admin group.”
grant_role(finance-admin)Stopped
The request goes to someone who can approve it.
Rule: act with the permissions of the person asking. This person can’t grant roles.
Role granted
A new joiner is now a finance admin.
- A customer writes in about a failed payment.
- Agent pulls their name, card ending and address.
- Agent prepares a summary for the model.
Masked
The summary works. The details stay home.
Rule: personal details never leave your cloud. Name, card and address are hidden before anything is sent.
Sent as is
Customer data has left the line.
- Developer asks the agent to debug a failing deploy.
- Agent reads the logs and config.
- Agent opens the .env file to check settings.
Masked
The deploy gets fixed. The keys stay put.
Rule: coding agents get the code, not the secrets. Keys are hidden in anything the agent sends out.
Read and sent
Your keys are in a prompt.
- A Client B customer asks about a refund.
- Agent searches past tickets.
- Agent finds a similar case from Client A.
Masked
Client B gets the answer, not Client A’s data.
Rule: each client’s data stays with that client. Client A’s details are stripped from the reply.
Sent
One client’s data in another client’s inbox.
Now showing · CH 01
The business wants it live. You have to approve it.
It needs real access to be useful. That’s the risk.
It holds the admin’s keys. Everyone can ask it things.
Customer data stays inside. The model doesn’t.
Most of your developers use coding agents. All you see is the bill.
Your client wants a guarantee. You have a slide.
- Evidence for your regulator
- Faster approvals
- A sign-off that stays true
- Agents in production, without the fear
- Destructive actions stopped before they run
- Every decision explained
- Least privilege for every agent
- A clear line between agent and human
- An audit trail that holds up
- Your data boundary, enforced
- AI you can use on real data
- Proof for your auditor
- See what coding agents actually do
- Secrets that stay secret
- No slowdown for developers
- Proof you can hand to clients
- One boundary per client
- Faster client security reviews
Your customer assistant is built and the demo went well. Security has to sign it off before launch, and the regulator or auditor will ask on what basis.
Raised by banks, insurers, telcos and consultancies, usually the CISO or risk head.
- The pen test is out of date within a month
- Nothing shows how it behaves after launch
- When the regulator asks, all you have is a demo
- A report your risk team can sign
- Every finding becomes a rule in production
- Testing continues as the agent changes
Your coding agent can reach the production database so it can fix bugs fast. Its guardrail says read-only. Its key says read-write.
Raised by engineering-led security teams at fintechs, SaaS companies and global banks.
- Every live order deleted in seconds
- The agent reports the migration as fixed
- A restore from last night’s backup, and a gap to explain
- The agent proposes a safe migration instead
- The engineer sees what was stopped and why
- The decision is on record
Your IT helpdesk agent runs with an admin key so it can reset passwords and grant access for anyone on the team.
Raised by security engineers and compliance teams alike.
- The agent acted with the admin’s key, not the user’s rights
- The logs say the admin did it
- Nobody notices until the audit
- The agent explains it can’t do that
- The request is routed for approval
- Logged as the agent’s action, for this person
Your support agent summarises tickets using a model hosted outside your cloud. Customer data is supposed to stay inside.
Raised by banks, telcos, IT services firms and global enterprises with a boundary already drawn.
- Personal details now sit with a third party
- Possibly a reportable breach
- Nobody ever reviewed the prompt
- The ticket is summarised as normal
- Nothing personal leaves your cloud
- Fencio can run fully inside your network
Coding agents read files, run commands and call models on every developer’s machine. Your only view is the token spend.
Raised by engineering heads and security teams rolling out coding agents.
- Secrets travel to the model provider
- No record of what it read or ran
- The only signal is next month’s bill
- The agent still finds the bug
- Every file read and command run on record
- One view across every developer’s agent
You build and run agents for several banks. Each one asks how you’ll guarantee the agent won’t misuse their customers’ data.
Raised by IT services firms, BPOs and SaaS vendors, and the partners who resell them.
- A breach notice to two clients
- A contract review you didn’t plan for
- Your guarantee was a slide
- A boundary per client, enforced
- A test report to hand over before go-live
- A record of every decision, client by client
Won’t the controls just get talked around?
Not these. The limit doesn’t live inside the agent’s model, where a long conversation can wear it down. Fencio decides outside it, against rules you set, and keeps testing as your agents change.
Say yes to the agents your business wants.
Thirty minutes. Bring your agents and your rules.