Tested Requests
0
Allowed (Compliant)
0
Blocked / Egress Dropped
0
Quarantine / Intercepted
0

Live Sandbox Egress Gateway Simulation

Policy Active: Enforcing
eBPF egress filter active. Inspecting agent tool invocations, subprocess curls, and API calls.
0 ms latency overhead
Ready to test custom agent action. Enter a URL or script call above.

Simulated Agent Actions (Training & Eval Batch)

Status Target & Method Agent Context / Tool Action Matched Rule Risk Tier

Why Internet Access Controls Matter During Training

During self-play, RL evaluation, and agentic training loops, foundation models generate shell commands, Python scripts, and tool calls. Without strict kernel-level egress filtering (via eBPF, Squid, or network namespaces), agents can:

  • Accidentally scrape proprietary/licensed datasets or download unverified binary weights.
  • Exfiltrate API credentials, environment variables, or private git tokens via webhooks.
  • Query local AWS/GCP instance metadata endpoints (169.254.169.254) to compromise IAM roles.
  • Pollute evaluation benchmarks by querying live test solutions from public forums.

Defense-in-Depth Architecture

This sandbox simulates an egress interception layer using three complementary mechanisms:

  • Domain & CIDR Whitelisting: Rejecting all raw IP requests and enforcing strict FQDN regex matching.
  • HTTP Verb & Protocol Pinning: Enforcing read-only GET/HEAD requests to prevent remote POST exfiltration.
  • Deep Payload Heuristics: Inspecting HTTP headers and outbound JSON for high-entropy secrets and credentials before packets leave the training cluster.
Enjoy this tool? Build your own with Super