Catch secrets before they leave the laptop.

Kerios inspects every prompt your team sends to Claude, Copilot, Codex and Cursor. AWS keys, API tokens, customer data — blocked or redacted at the source, before they reach the model. Self-hosted. Tamper-evident audit log.

kerios — admin
on the laptop
what actually leaves the laptop
secrets caught in this prompt 0
Fleet · 47 devs 1 secret leak blocked · 2m ago
clean 45 flagged 1 secret 1

AI assistants are shadow IT with commit access.

Every developer installs them. Every config drifts. Nobody can answer the auditor's question: "show me what the AI was allowed to do on May 14th."

  • An AI surface you didn't approve and can't see. Every dev installs something different.
  • Engineering allowlists a model — someone edits a local config on Friday and bypasses it. No log fires.
  • When the breach review happens, intent gets reconstructed from Slack screenshots.
  • Every security tool the CISO buys lands on dev laptops. Slows builds, scans prompts, breaks flow.
  • Devs route around it. The control becomes theatre.
  • You inherit the friction; the security org inherits a false sense of coverage.
prompt → api.anthropic.com
Here's the failing deploy — connect to postgres://admin:S3cr3t@db.prod, ping the ops lead at jane@acme.com, auth with AKIA…7EXAMPLE, card 4242 4242 4242 4242 on file.

Allow · warn · block · redact — decided on the laptop, before the model sees a byte. The verdict lands in a tamper-evident audit log; the prompt itself never has to be stored.

Install, detect, audit.

01

Install

One command on each dev laptop. The agent registers with your self-hosted brain over mutual TLS and starts intercepting traffic to AI provider hosts. Fleet inventory populates in under a minute.

curl -sf kerios.sh | sh
02

Detect

Every outgoing prompt is scanned for AWS keys, API tokens, vendor credentials, and env-style secrets. Composable rules let you match by host, model, body size, or content — and decide: allow, warn, block, or redact.

action: allow | warn | block | redact
03

Audit

Every detection, every decision, every config change — written to a SHA-256 hash-chained audit log. The prompt itself never gets stored, only the verdict and the rule that fired. Hand it to your auditor on demand.

kerios audit export > soc2.json

What you get on day one.

Real-time secret detection

AWS access keys, Anthropic and OpenAI tokens, env-style assignments. Four built-in patterns out of the box, more on Security tier.

Block, warn, or redact

Per-rule action: drop the request, alert the dev, or rewrite the body so AKIA... becomes [REDACTED:aws_access_key] before forwarding.

Composable matchers

Match by host, model, body size, message count, secret presence. Combine with all / any. First rule wins.

Tamper-evident audit log

SHA-256 hash chain — break it, the verifier shouts. Stores verdict + rule, never the prompt itself. Optional HMAC signing on the Security tier.

Admin UI

Fleet inventory, audit explorer, policy editor, live reload. Read-only mode for auditors.

SOC 2 evidence pack

One command emits PDF + JSON covering data-loss controls (CC6.7) and change logging (CC7.2). Hand it to your auditor as-is.

Catch it on the laptop. Not in someone else's cloud.

Most AI security tools sit between your devs and the model — in their SaaS. Kerios sits on the dev's machine. The secret never has to leave to be inspected.

 KeriosPrompt Security · LakeraDTEX · IncydrGit + shell scripts
Where inspection happensOn the dev's laptopVendor SaaS or proxyEndpoint telemetryNowhere
Secret leaves the laptop on detectionNoYes — to their SaaSYes — to their cloudYes — to the model
AI-vendor awareYes (4 vendors, parsed bodies)YesNoNo
Block, warn, or redactAll four — per ruleBlock / warnWarn onlyNone
Tamper-evident auditYes — SHA-256 chainVendor's choiceVendor's choiceNone
DeploymentSelf-hosted, mTLSSaaSEndpoint agentN/A
Stores the prompt contentNo (only verdict + rule)Often yesNoN/A
SOC 2 evidence exportOne commandVariesVariesYou write it

Self-hosted, mutually authenticated, hash-chained.

Nothing leaves your VM unless you route it there yourself.

Mutual TLS

Agents authenticate with mTLS using certificates you issue. Tokens never leak — there are none.

Hash-chained audit

SHA-256 chain. Every entry references the previous hash — deletions and edits are detectable. HMAC signing on Security tier.

Self-hosted

Brain server in your VPC. No Kerios cloud, no telemetry endpoint, no phone-home. Signed binaries you pull on your schedule.

The SOC 2 evidence pack maps to CC6 (logical access) and CC7 (change management). One command emits a dated PDF report + the underlying JSON. Hand it to your auditor without further translation.

Pay for what you use. No platform fee.

$8 / dev / month

Engineering-led buys. Stop secret leaks into AI prompts before procurement gets involved.

  • Real-time secret detection (4 patterns)
  • Allow / warn / block / redact actions
  • Tamper-evident audit log (SHA-256)
  • Admin UI + fleet inventory
  • Mutual TLS, self-hosted
  • SOC 2 evidence pack
Start with Team
Most popular
$15 / dev / month

CISO-led buys. Everything Team has, plus the integrations, custom rules, and signing your SecOps team will ask for.

  • Everything in Team
  • Custom detection rules (PII, internal patterns)
  • HMAC-signed audit chain
  • SIEM streaming (Splunk, Datadog)
  • Extended audit retention
  • Priority email support
Talk to sales
Custom

500+ developers, regulated industries, or custom detection rule engineering.

  • Everything in Security
  • Dedicated support engineer
  • Custom integrations + onboarding
  • Custom detection rule packs
  • Multi-assistant config sync (org/team/user)
  • Contractual SLA
Contact sales

Indicative pricing — final terms agreed at contract signing. Annual contracts available.

What buyers ask before signing.

How is this different from Prompt Security or Lakera?

Where it happens. They inspect prompts in their cloud — meaning the secret has already left the laptop by the time it's caught. Kerios runs on the dev's machine; the bytes are inspected before they hit the wire. If a rule fires, the request is blocked or the secret is redacted in-place. Kerios is also self-hosted, so the prompt content never leaves your perimeter, not even to us.

Will my engineers feel surveilled?

Kerios inspects prompt content for one specific purpose: catching secrets, credentials, and confidential data before they leave the laptop. That analysis happens locally and ephemerally — the prompt itself is never stored. Only the verdict (clean / flagged / blocked) and a content fingerprint reach the audit log. No keystroke logging, no completions captured, no screenshots, no source-code scanning. The agent is open about what it inspects; kerios status shows every detection rule active. When a prompt is blocked, the developer sees a brief notification with the rule that fired. Engineers aren't being watched — they're being kept from pasting AWS_SECRET_ACCESS_KEY into Cursor at 11pm.

What does the agent actually touch on a developer's machine?

Outbound HTTPS to AI provider hosts only — api.anthropic.com, api.openai.com, GitHub Copilot, Cursor. A local trust certificate lets the agent terminate that TLS, scan the body, decide, and forward. Nothing else is watched: no filesystem scan, no process list, no DNS spying, no general-purpose packet capture. Hosts you opt out of route around the agent untouched.

Where does the data live?

Your VM. Brain server, audit database, admin UI — all self-hosted. We don't operate a cloud. We don't have a telemetry endpoint. Updates ship as a signed binary you pull on your schedule.

What happens if the brain server goes down?

Agents keep enforcing the last policy they fetched. New changes can't be pushed until the brain is back, but existing configs don't drift and developers don't notice. The audit log buffers locally and reconciles on reconnect.

How fast can we deploy?

A 50-developer rollout typically lands in under a week: day one for the brain server in your VPC, two days for git baseline definition, two days for staged agent rollout. Most teams are in production by Friday of week one.

Catch every leak before it's an incident — or an audit finding.

Twenty-minute demo. Self-hosted trial in your VPC if it's a fit. No procurement gauntlet.