Join early access

Elsewhere X @myriaxlabs (opens in a new tab)

Defense · planned, early access

Stop rogue agents.

Block the agents attacking your site and API, and keep the agents you run inside their limits. Defense isn't live yet.

Myriax Defense · planned

Two sides of the same problem.

Agents now hit your API from both directions. Defense will block the ones attacking you and limit the ones you run.

At your door

Block hostile agents

Built to stop exploit bots and spoofed crawlers before they reach your site or API.

Request log Sample
  1. block GET /.env scanner probe
  2. allow GET /docs signed agent
  3. limit POST /v1/chat spend limit
  4. block GET /fetch?u=169.254.169.254 SSRF to metadata
  5. observe GET /../../etc/passwd new rule
block
2
limit
1
allow
1
observe
1
Exploit rules
Rules for path traversal, SSRF to metadata and scanner bursts.
Bot identity
Verified crawlers and signed agents (Web Bot Auth) will be allowed; spoofed ones will be blocked.
Rate and spend limits
Per-identity limits on your AI endpoints.
Observe first
Each new rule will start observe-only, so you see it before it blocks.

Illustrative. Rules recognise attacks; they don't patch your app, so keep patching.

Inside your stack

Keep your agents in bounds

Your own agents will get permissions, a budget and a switch that stops them.

agent: support-bot Running
Kill switch
tools
approval issue_refund needs a human
budget $3.80 of $10.00
egress api.helpdesk.example
Tool permissions
Deny by default, with argument checks and approval hooks.
Hard spend limits
Budgets will be reserved before each call, so replicas can't overspend.
Kill switch
Will stop one agent or all, through the gateway, on the next request.
Audit log
Will log every denial with its reason and policy version, secrets redacted.

Illustrative. Planned: enforced at the Myriax gateway for routed calls, and in the SDK for wrapped tools.

How it works · planned

  • Built to check every request before it reaches your API or your agents' tools, adding as little latency as possible.
  • Designed to block only what's clearly malicious, and challenge or hold what's unclear.
  • You'll stay in control: spend limits, approvals and a kill switch. Nothing blocks without rules you've approved.

Why now: exploiting a vulnerability was the way in for 31% of breaches in Verizon's 2026 DBIR (SC World), and CISA added 245 actively exploited flaws to its KEV catalog in 2025 (SecurityWeek).

Sample conversation

Ask Defense what it did, and why.

The plan is for Defense to explain in plain words what your rules blocked and why. The agent would explain; it wouldn't hold the block switch. Pick a question to see how a conversation could go.

Defense agent Illustrative sample · Defense isn't live

Pick a question below.

Early access

Get in before launch.

Early access opens to a small group of teams first. Tell us what you'd run through Myriax.

Join early access