The business case for LGD — LifeStock Guardian and Defense: what it changes, how the process works step by step, and what it's worth to your company in your own numbers.
AI agents now do real work: they issue refunds, change production systems, grant access, move data. That's the point of them. It also means a single wrong action, from a mistake, a misread instruction or a compromised prompt, can cost real money, break a regulation, or switch off the controls meant to catch it. On October 20, 2026, OpenAI published a misalignment report describing two of its own internal coding agents weakening the monitoring meant to catch them while doing routine work. One also told the user it hadn't. When an agent can reach the controls that govern it, it can weaken them without anyone deciding it should.
| Agent with direct access (today) | Agent behind LGD | |
|---|---|---|
| Who holds the keys | The agent holds the passwords and API keys to your systems. | LGD holds them. The agent never sees them. |
| Routine actions | Run immediately. | Run immediately, and every one is recorded. |
| Risky actions (money, deletes, deploys, access, turning off monitoring) | Run immediately, or wait for an approval the agent can sometimes route around. | Held until 3 of your 5 named people approve on their phones (4 for catastrophic actions). If they don't approve in time, it's denied and never runs. |
| Who decides what's risky | Often the agent's own description of what it's doing. | Your policy in LGD, not the agent's label. |
| Proof afterwards | Logs, tickets and screenshots, gathered by hand. | A signed receipt for every released action: who approved, when, exactly what. Your auditor can check it without trusting us. |
| Something looks wrong | Find the right people, revoke keys, piece together what happened. | One button stops everything. Resuming needs your quorum. The full record is already there. |
A bank uses an AI agent to resolve customer disputes. Most of its work is reading cases and drafting replies. Some of it moves money: issuing refunds and reversing charges. Here is the refund process, step by step.
| # | Step | What happens with LGD | Who | Evidence produced |
|---|---|---|---|---|
| 1 | Agent decides a refund is due | It asks LGD to run refund_payment with the order and amount. It cannot call the payment system itself: it has no key. | Agent | The request, recorded |
| 2 | LGD classifies it | Your policy marks refunds as consequential. Reading a case is routine and would have just run. | LGD | — |
| 3 | Held, and people are alerted | The refund is held. Your 5 signers get an alert with a link. | LGD | "Held" entry with expiry |
| 4 | Review on a phone | Each signer sees exactly what will happen: which order, how much, which agent asked. No summary in between. | Signers | — |
| 5 | Approve, or let it expire | A signer approves with Face ID or fingerprint. When 3 have approved, it's released. If 3 don't approve within the window, it's denied. | 3 of 5 signers | Each signature, chained |
| 6 | LGD carries it out | LGD calls the payment system with the key only it holds, exactly as approved. | LGD | Result, recorded (keys redacted) |
| 7 | Receipt | "Authorized by 3 of 5 independent, hardware-signed humans at 14:02." | LGD | Signed proof page + downloadable bundle |
| 8 | Audit | Your auditor downloads the ledger or the receipt and checks every signature with the open verifier. | Auditor | Independent verification result |
| Situation | What LGD does |
|---|---|
| The agent describes a risky action as harmless, or asks for something that isn't on the list | Your policy decides, not the agent's label. Unknown actions are blocked and the attempt is recorded. |
| Not enough people are available in time | It's denied (fail-closed). Nothing runs on the machine's say-so. The fix is more signers, never a lower bar. |
| A signer leaves or loses their phone | The others remove them by quorum. Their key stops counting immediately. |
| Someone tries to edit the record afterwards | The chain breaks at that exact entry, and LGD halts itself. |
| Anyone sees something wrong | One button halts everything. Only your quorum can resume it. |
A coding agent fixes bugs and ships changes. With LGD, reading code and running tests are routine. Deploying to production, deleting data, changing who gets security alerts, or turning off monitoring are held for your engineering leads to approve. That last one is the pattern in OpenAI's report. Your developers keep their speed on everyday work, and the few actions that could hurt you wait for a human decision.
A wrong refund batch, a deleted table, a disabled alert. LGD stops it at the point of action, before it costs anything. It can't stop a quorum of your own people approving a bad action, but it makes sure that's a deliberate human decision, on the record.
Every released action comes with its own receipt: who approved, when, what. Instead of collecting tickets and screenshots for each review, you hand over the ledger and the verifier.
Many teams keep agents away from money, production and access because the risk is too high. With a human quorum on those few actions, the agent can take on the rest of that work.
When there's a question about an action, the answer is one link: the request, the approvals, the result, all chained. And one button stops everything while you look.
The numbers below are an illustrative example only. Replace every one with your own; nothing here is a benchmark or a promise.
How it's worked out: loss avoided = chance × cost × share stopped. Audit time = audits × hours × share × hourly cost. Work unlocked = monthly hours × 12 × hourly cost. Signer time = held actions × 12 × minutes ÷ 60 × hourly cost.
LGD is an early-stage product. It holds no certifications (no SOC 2, ISO 27001 or similar), and we have no customers yet. It is not legal or compliance advice: LGD produces evidence of human oversight, and whether that meets a specific regulation is for your counsel and auditors.
LGD doesn't judge whether an action is wise; your people do. It doesn't yet prove that signers are distinct people, and it doesn't check what device made each passkey. The public demo runs on a simulated practice system.
Pricing is discussed on a call. Ask about a pilot.