Private beta · invite-only

AI traffic doesn't belong on a SaaS API.
It belongs on your firewall.

Kilasec is an AI-aware firewall for enterprise networks. Discover every model API call, redact secrets and PII before they leave, and cap runaway agent spend — without an SDK, and without installing anything on user devices.

$ Drops in as a network appliance. One config push to your existing proxy/PAC. No endpoint agent, no SDK.
Built into the network

Every other AI security tool is an SDK.
This one is a firewall.

The current crop of AI security startups asks you to integrate a library, run a sidecar, or rewrite your agents. That works exactly until someone in marketing pastes an SSN into ChatGPT, or a vendor's tool you've never heard of starts calling Anthropic from a finance laptop. Kilasec sits in the network where you already enforce policy.

See every AI call

Decrypted inspection of OpenAI, Anthropic, Copilot, Gemini, Bedrock, Ollama, and 30+ other endpoints — including ones we haven't named yet, classified on the fly.

Redact at the boundary

Secrets, credentials, customer PII, and credit cards are caught and masked the moment they leave your network — not when an SDK says please.

Identity that fits your stack

Map every request to a real user and AD group — from your directory, your DHCP server, or a CSV upload. Write rules like allow group:engineering to api.anthropic.com instead of staring at IPs.

Audit-grade history

Every admin change and every blocked request, with configurable retention from 30 days to 7 years. Export for SOC2, HIPAA, or your own change-management process.

vs the alternatives

Why network-layer beats SDK-layer

SDK approaches assume you control every code path that calls an LLM. You don't. Browser extensions, third-party tools, agents your engineers downloaded last week — none of them link your library.

SDK / library
Lakera, Straiker, PolicyLayer
URL filtering / SWG
Netskope, Zscaler
Kilasec
Network-layer firewall
Catches AI tools you didn't know about No — only what you instrument Partial — domain only Yes — every TLS flow
Redacts secrets & PII in prompts Yes No — payload not inspected Yes
Identity per request Library-supplied (spoofable) User-Agent / SSO User + AD group (network-resolved)
Token / cost visibility per agent Only instrumented agents No Yes — every model, every call
Deployment effort Code change in every app Endpoint agent rollout One push via your network
Coverage of unmanaged BYOD & vendors No If on managed device Yes — anything on the network
How it works

One collector. One config push. Every AI call.

The collector is a single Linux container running a TLS-decrypting proxy and our policy engine. Your existing network advertises it to every device — no install on user laptops, no SDK in your apps.

Laptop · Agent · VM unmodified no SDK installed Your network pushes proxy config boot config Collector TLS-decrypting proxy on-prem · 1 container AI traffic only Policy engine allow · deny · redact · approve kilasec.com admin UI · reverse tunnel WSS api.openai.com api.anthropic.com copilot · gemini · … unapproved → blocked 1 2 3 4
1. Device boots. Picks up its network config — and the proxy settings — automatically.
2. AI traffic flows. Only AI hosts go through the collector. Everything else is direct.
3. Policy decides. First-match-wins rules: allow, deny, redact, or require approval.
4. You see it all. The collector streams decisions to the cloud UI over a reverse tunnel.
What it catches

The threats actually showing up in our beta networks

None of these are hypothetical. They're the events our policy engine flagged in real customer environments during the closed beta.

Secret leak

AWS credentials pasted into ChatGPT

An engineer asked ChatGPT to help debug a script. The script had hardcoded AWS_ACCESS_KEY and AWS_SECRET. Both would have left the network.

Blocked at the proxy, before TLS unwrap on the OpenAI side
PII redaction

Customer SSNs sent to support agent

A support team's AI agent received raw chat transcripts containing customer SSNs and credit card numbers. None of it should have reached Anthropic.

Replaced with placeholders inline, original kept on-prem
Shadow AI

Unknown vendor calling DeepSeek

A new vendor tool started calling api.deepseek.com from a finance laptop. Nobody in IT had ever heard of it.

Surfaced on first request — provider not on the allowlist
Cost runaway

Eval loop burned $300 in 18 minutes

A misconfigured QA agent started rerunning evals on Opus. At that rate, $50K by morning. The daily-spend rule paused it.

Held for approval at $50, owner pinged
The console

A firewall view, not a dashboard view

Designed for the network operator, not the AI engineer. Ordered rules with first-match-wins, dense traffic logs, real user identity, a tabbed rule editor that reads like Palo Alto — not a Slackbot configurator.

kilasec.com/app/decisions
Overview
Dashboard
Incidents
Security
Live Traffic
Agents
Policy Rules
Certificates
Operations
Destinations
Approvals
Traffic Log
Audit Log
Workspace
Collectors
Identity
Settings
AI requests
142,318
↑ 12.4% vs yesterday
Blocked
847
PII · secrets · shadow AI
Sensitive caught
312
SSN · keys · credentials
Spend today
$1,284
51% of $2,500 cap
Time
Verdict
Agent
Destination
Reason
14:02:47.218
DENY
openai-python
api.openai.com/v1/chat
aws_key
14:02:46.842
REDACT
anthropic-python
api.anthropic.com/v1/messages
ssn ×3
14:02:46.318
DENY
browser
api.deepseek.com/v1/chat
banned
14:02:46.014
ALLOW
claude-cli
api.anthropic.com/v1/messages
group:engineering
14:02:45.612
APPROVE
cursor
api.anthropic.com/v1/messages
cost > $200/d
14:02:45.401
ALLOW
github-copilot
api.githubcopilot.com
approved
Shipped in v0.5

What's new this release

The product is iterating fast. Highlights from the last few weeks of beta feedback.

Tabbed rule editor

Rule edit reorganised into General / Source / Destination / Schedule / Action tabs with a live plain-English preview at the top. Chip-style inputs for users, groups, hosts, tools, models. One-click presets to add all OpenAI or Anthropic endpoints at once.

User-ID for AI

Map source IPs to real users and AD groups from your directory, your DHCP server, or a CSV. Rules can now match on user: or group: instead of raw IPs — the PA User-ID model, applied to AI traffic.

Real Audit Log

Separate from the Traffic Log: every admin change — policy edit, collector approval, retention change, identity mapping — captured with actor, timestamp, target, and JSON diff. CSV export. Filter by actor / action / date.

Active-hours rules

Restrict rules to specific hours and days of the week with timezone awareness. Block agent calls after 17:00 PT Mon–Fri is now a one-click chip toggle, not a regex.

Destinations catalog

Admin-curated address objects with labels and categories (approved / experimental / banned). Bulk-import from CSV. The Editor's host chip autocomplete suggests both observed and registered destinations.

Agent identification

Every TLS-decrypted request now resolves to a normalised agent name from the User-Agent header — claude-cli, openai-python, cursor, github-copilot, browser — instead of one giant unknown_agent bucket.

Get on the beta

We onboard 1–2 networks a week. Send us your work email and we'll set up your collector within 24 hours of approval.

No credit card. Free during beta.