01Borrowed authority
The assistant inherits access nobody granted it.
It was wired up with a service credential so it would 'just work'. Now it reads every row a human never could, and no ticket records the moment that started.
Deep-tech security · Est. 2026
Give autonomy
a chain of command.
Software that acts on its own is already inside your stack — booking payments, reading customer records, calling tools you never reviewed. CuriousDevs puts a decision checkpoint in front of every one of those actions, so autonomy stays answerable.
Built as one stack
01AgentGuard
02CurioComply
03AeroOS
<10ms
decision time
No model
in the hot path
Signed
every outcome
Why now2026 — 2030
45:1Non-human to human identitiestypical enterprise estate
13 May 2027India's data-protection cutovera date nobody can move
₹250 crMaximum penalty per violationunder the DPDP Act
6 hoursIncident reporting windowCERT-In directions
<10msVerdict on every tool callno model in the hot path
3 layersAgents, data, machinesone accountability spine
45:1Non-human to human identitiestypical enterprise estate
13 May 2027India's data-protection cutovera date nobody can move
₹250 crMaximum penalty per violationunder the DPDP Act
6 hoursIncident reporting windowCERT-In directions
<10msVerdict on every tool callno model in the hot path
3 layersAgents, data, machinesone accountability spine
45:1Non-human to human identitiestypical enterprise estate
13 May 2027India's data-protection cutovera date nobody can move
₹250 crMaximum penalty per violationunder the DPDP Act
6 hoursIncident reporting windowCERT-In directions
<10msVerdict on every tool callno model in the hot path
3 layersAgents, data, machinesone accountability spine
45:1Non-human to human identitiestypical enterprise estate
13 May 2027India's data-protection cutovera date nobody can move
₹250 crMaximum penalty per violationunder the DPDP Act
6 hoursIncident reporting windowCERT-In directions
<10msVerdict on every tool callno model in the hot path
3 layersAgents, data, machinesone accountability spine
Where it breaks
Nothing alarms when autonomy goes wrong.
There is no stack trace for a decision. The system runs green while doing precisely the wrong thing, and you find out from a customer, a regulator or a bank statement.
02Instruction hijack
A document tells the agent what to do next.
The page it retrieved carries a sentence written for the model, not the reader. The agent obeys. Your prompt never mentioned any of it.
03Quiet data exit
Personal records leave through a helper call.
One integration widens its payload during an upgrade. Identity fields ride along to a vendor you never mapped, and the export looks routine in the logs.
04Machines in motion
Nothing broke. The robot simply improvised.
The task was under-specified, so the fleet resolved the ambiguity itself. Every individual choice was defensible; the combined outcome was not.
The missing control
Everything you own today watches, none of it decides.
Four familiar categories, one shared blind spot: none of them stand between intent and execution while the system is live.
Dashboards
Recorded
They narrate the incident after the money left. Useful for the write-up, powerless at the moment of action.
Prompt rules
Suggested
Text inside a prompt is persuasion. A long context, a clever input or a bad day is enough to talk the model out of it.
Annual audits
Snapshot
A binder describing last quarter's system. Your data flows changed on Tuesday and the evidence never caught up.
Model reviewers
Approximate
Asking a model to police a model doubles both the uncertainty and the bill, and still yields no reproducible verdict.
How it Works
From failure modes to runtime enforcement.
Pre-Production · Control Plane
Discovery & Evals
Adversarial · Security · Compliance
QualitySecurityDPDP
Sandboxed reconstruction
Behavioural Baseline
Risk & compliance scoring
Compliance RiskSecurity Score
From observed failure modes
Policy Signatures
Reusable rules, deterministic, not prompt-based
Data ProtectionGovernanceContext Mgmt
Shift-left risk discovery
↓ Unified Policy Engine · Deploy
Runtime · Data Plane
Policy Enforcement
Detect · Allow · Deny · before execution
AllowDenyModifyEscalate
Inline, deterministic, real-time
Real-time Monitoring
Every signal observed continuously
User InputLLM CallsTool Calls
Full audit trail on every decision
Continuous Hardening
Production tightens enforcement
TelemetryPolicy Updates
No redeployment. No system modification.
01 · Pre-Production
Map & Evaluate
Data maps and adversarial evals run against a sandboxed reconstruction of your agents and pipelines.
02 · Pre-Production
Build Baseline
Behavioural baseline established. Risk and compliance scores generated from observed failure modes.
03 · Pre-Production
Generate Policies
Failure traces become deterministic policy signatures, reusable across every agent and fleet.
04 · Runtime
Enforce & Monitor
Every tool call, LLM call, user input and machine command passes through the enforcement layer in real time.
05 · The Loop
Prove & Harden
Evidence exports map to SOC 2, ISO 27001 and DPDP. Telemetry feeds back and enforcement tightens continuously.
The path, 2026 to 2030
Ambitious, but in order.
2026
AgentGuard goes public, core in the open
Gateway, SDKs and the rule engine ship where engineers can read them. Trust in security software is earned by being inspectable.
2026–27
CurioComply opens with a free scan
Anyone can see their exposure in two minutes. Teams that like the answer stay for the automation underneath it.
2027
The compliance cutover
India's obligations become enforceable. We intend to be the boring, obvious choice by the time procurement starts calling.
2027–28
AeroOS leaves the lab
Mission command for mixed fleets, standing on the checkpoint and evidence layers already in production.
2030
Where this is going
When a board asks how they can trust machines acting without a human in the loop, we want the answer to be a product, not a policy memo.
What we believe
“Any action a machine takes on its own should have a name attached, a reason recorded, and someone who can answer for it.”
- 01Security is the whole product, never a module bolted on later.
- 02Build from India, measure against the best anywhere.
- 03Open the core; developers adopt what they can read.
- 04Treat regulation as demand, not as drag.
- 05Every layer must make the next one easier to sell.
Let something decide for you. Just not unsupervised.
Bring one workflow you are nervous about. We will show you exactly where the checkpoint goes and what it would have stopped last month.
Questions
The things people ask first
Short answers on how this runs, what it costs you in speed, and what you can hand to an auditor.
A decision checkpoint for software and machines that act on their own — it approves, trims, escalates or refuses each action, and keeps a signed record of why.