Skip to content

Releases: offenselabs/EgressEye

Release list

EgressEye v0.1.0 — first public release

EgressEye v0.1.0 — first public release Latest
Latest

Choose a tag to compare

View all tags
@0xSpecters 0xSpecters released this 29 Oct 23:04
v0.1.0
b4b5e9f

First public release of EgressEye — per-session egress visibility for AI coding agents.

What it does

Sits in your harness's hook layer, inspects every tool call, and answers one question: what did this session contact, what looked wrong, and why?

  • Deterministic detection — OpenPhish/URLhaus mirror feeds (local, in-memory), optional key-gated Google Safe Browsing, typosquat/punycode/IP-literal/credentials-in-URL heuristics against a dev-domain allowlist, RDAP newly-registered-domain checks. No LLM in the verdict path.
  • Secret scanning — AWS/GitHub/OpenAI/Anthropic/Slack/Google tokens, JWTs, private keys, DB connection strings, .env blobs; redacted at creation (4-char preview + length only).
  • Sharded-exfiltration aggregation — many small sends to the same host (path rotation included) surface as one session-level finding.
  • Honest verdicts — unknown means unknown; every verdict carries a machine source and human-readable reason.
  • Local-first privacy — URLs stored query-stripped, secrets redacted on write, feed lookups in memory, zero runtime dependencies (Node builtins only).

Harness adapters (all driven by the same engine)

Harness Install
ZCode egress-eye init zcode
Claude Code egress-eye init claude-code
Codex egress-eye init codex
Hermes (Nous Research) egress-eye init hermes
OpenCode egress-eye init opencode

DeepSeek Harness (dsh) is researched and feasible — deferred until its plugin API stabilizes (see docs/adapters/deepseek-feasibility.md).

Getting started

npm install                      # build (TypeScript dev deps only)
node dist/cli.js init zcode      # pick your harness
node dist/cli.js update-feeds    # mirror the blocklist once
# … use your agent normally …
node dist/cli.js report --last

dist.zip below contains the compiled tool if you don't want to build.

Verification

  • 70/70 unit tests (node --test, fully offline)
  • 39/39 eval-scenario corpus replayed through the real hook path
  • CI green on ubuntu + windows, Node 20 and 22

Honest limits

EgressEye is a visibility tool, not a sandbox. The complete bypass list is in docs/limits.md and the threat model in docs/threat-model.md — both are part of the product.

Commits

  • egress-eye v0.1.0: per-session egress visibility for AI coding agents
  • adapters: Claude Code, Codex, Hermes (Nous Research), OpenCode + dsh feasibility
  • audit fixes: independent review findings (5 important + minors)
  • chore(release): beginner-friendly README, CI workflow, repo metadata
  • test: avoid real-format token literals so push protection stays enabled
  • fix(ci): drop trailing slash from the test runner path
  • fix(ci): enumerate test files explicitly for cross-version node --test

Full Changelog: initial release

Assets 3
Loading