Durable research for building a macOS host-safety layer that protects a dev box from
any AI agent or downloaded executable running on it — the primitives, prior art, the
read-only-vs-privileged trust map, and the no-keyword-NLP secret-detection method behind
Port Daddy's `pd safe` ("the AI little sniffer", ADR-0088). Covers Endpoint Security
Framework (ESF) + eslogger, Network Extension (NEFilterDataProvider/DNS/transparent
proxy), pf forced egress, code-trust CLIs (codesign/spctl/xattr), libproc/nettop/lsof
read-only inventory, TCC + Full Disk Access, structured-format + Shannon-entropy secret
scanning, and Santa-fronted binary lockdown. Activate on: "macOS host security", "endpoint
security framework", "eslogger", "ESF AUTH_EXEC", "NEFilterDataProvider", "network
extension", "system extension entitlement", "pf forced egress", "codesign trust",
"quarantine xattr", "Full Disk Access", "TCC", "secret scanner", "gitleaks rules",
"Shannon entropy detection", "Santa binary lockdown", "pd safe", "AI sniffer", "con
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
license
MIT
name
macos-host-security
description
Durable research for building a macOS host-safety layer that protects a dev box from
any AI agent or downloaded executable running on it — the primitives, prior art, the
read-only-vs-privileged trust map, and the no-keyword-NLP secret-detection method behind
Port Daddy's `pd safe` ("the AI little sniffer", ADR-0088). Covers Endpoint Security
Framework (ESF) + eslogger, Network Extension (NEFilterDataProvider/DNS/transparent
proxy), pf forced egress, code-trust CLIs (codesign/spctl/xattr), libproc/nettop/lsof
read-only inventory, TCC + Full Disk Access, structured-format + Shannon-entropy secret
scanning, and Santa-fronted binary lockdown. Activate on: "macOS host security", "endpoint
security framework", "eslogger", "ESF AUTH_EXEC", "NEFilterDataProvider", "network
extension", "system extension entitlement", "pf forced egress", "codesign trust",
"quarantine xattr", "Full Disk Access", "TCC", "secret scanner", "gitleaks rules",
"Shannon entropy detection", "Santa binary lockdown", "pd safe", "AI sniffer", "confine
agent on macOS", "block rogue executable". NOT for: the TypeScript daemon's coordination
logic, generic Rust/borrow-checker help (use rust-with-claude-code), web/UI security
(use agentic-zero-trust-security for crypto/ocap), or prompt-injection defense.
allowed-tools
Read,Write,Edit,Bash,Glob,Grep,WebSearch,WebFetch
metadata
{"category":"Security & Trust","tags":["macos-security","endpoint-security","eslogger","network-extension","system-extension","pf-firewall","codesign","tcc","secret-detection","shannon-entropy","santa","host-safety","pd-safe"],"pairs-with":[{"skill":"agentic-zero-trust-security","reason":"The cryptographic/ocap layer; this skill is the OS-primitive layer beneath it."},{"skill":"daemon-development","reason":"pd safe's read-only sensors run in the Port Daddy daemon."}]}
macOS Host Security — the pd safe research base
Durable record of the macOS host-safety research so future sessions don't re-research it.
This is the OS-primitive foundation under pd safe (the AI little sniffer,
ADR-0088). The product widens the Coast
Guard (lib/coast-guard.ts, ADR-0050) from protect git push to protect the whole
machine from any agent/executable.
The cardinal rule (read this first, every time)
Any tool you run as the operator's own UID, a same-UID agent can disable, unset,
kill, or read around. Read-only sensors are genuine DETECTION and EVIDENCE — they
catch runaway loops, accidental exfil, confused-deputy reads, a freshly-downloaded
unsigned binary phoning home. They are NOT containment. Real containment needs an
authority outside the agent's UID: a SIP-protected system extension (ESF or NE) or a
separate UID/VM + pf. Never market a same-UID watcher as a wall. lib/coast-guard.tsHONEST_LIMITS states this verbatim; echo it.
TCC trap: ~/.env/~/.aws/~/.ssh/~/.config/~/.npmrc are plain POSIX-readable by
the same UID — TCC does not guard them. Full Disk Access guards ~/Library/Mail| Messages|Safari and other apps' containers, not the dotfiles that hold the crown jewels.
Trust per binary:codesign --verify --deep --strict + codesign -dv --verbose=4
(Authority chain, TeamIdentifier, CDHash, ad-hoc flag) + codesign --check-notarization. Use SecCode / codesign for bare CLI binaries; reserve spctl -a -t exec for .app bundles only (it misreports standalone CLI executables — the
common agent/MCP shim shape). csrutil status for SIP, spctl --status for Gatekeeper.
Provenance:xattr -p com.apple.quarantine. Missing quarantine = UNKNOWN, never
SAFE — curl|bash, scp, git-clone, npm/pip leave no quarantine xattr; that is the
dangerous path. Lean on signing + notarization, not provenance.
Attribution: use ESF responsible_audit_token, NOT ppid — on macOS the parent is
almost always launchd/runningboardd/xpcproxy. A ps-only sniffer mis-blames launchd
for everything.
Scan the KnockKnock location set, diff against a baseline, alert on NEW items:
~/Library/LaunchAgents, /Library/LaunchAgents, /Library/LaunchDaemons plists,
login items (SMAppService / BackgroundItems .btm), cron (crontab -l, /etc/periodic),
shell rc files, launchctl print.
Egress observability (read-only)
nettop -P -m route -l 1 — per-PID host + byte counters, no sudo. Sampled, not
per-flow-complete (short flows between samples are missed).
lsof -i -nP — own-UID sockets (root for cross-process).
Destination without MITM: TLS SNI is cleartext after the handshake (RFC 6066) —
read the hostname without terminating TLS. Encrypted Client Hello (ECH, RFC 9849,
rolling out ~2026) erases SNI where deployed — fall back to IP + rDNS.
log stream --predicate 'subsystem == "com.apple.network" AND category == "connection"'.
Beaconing (RITA-style, explainable, NO ML/keywords): low coefficient-of-variation on
inter-request intervals + low time-gap entropy + byte-ratio spikes. Rank + explain; never
binary alarms (legit AI agents are bursty and many-hosted).
Real-time telemetry via eslogger (read-only, elevated)
eslogger exec fork open uipc_connect ships in macOS 13+ already entitled — ESF NOTIFY
telemetry with NO Apple entitlement; cost is sudo + the host terminal/app holding Full
Disk Access. NOTIFY-only by design — it cannot block. Output is explicitly "NOT API":
schema can change between macOS releases, so parse defensively (tolerate missing/renamed
fields) and version-pin. ESF emits NO general socket-connect events (only UNIX-domain
UIPC_CONNECT) — use ESF for process provenance, Network Extension for the flow. The
tstromberg/esl Go project is the reference for wrapping eslogger.
Real enforcement (privileged)
Endpoint Security AUTH:es_new_client → subscribe ES_EVENT_TYPE_AUTH_EXEC /
AUTH_OPEN → es_respond_auth_result(ALLOW|DENY)before es_message_t.deadline (an
absolute mach time — miss it and the kernel kills your client; a dead default-deny client
can soft-brick exec). Mandatory: verdict cache + es_mute_path/es_mute_process.
Requires the Apple-granted com.apple.developer.endpoint-security.client entitlement.
Network Extension:NEFilterDataProvider (per-flow allow/drop keyed on
flow.sourceAppAuditToken — survives unset HTTPS_PROXY), NEDNSProxyProvider
(per-process DNS names; DoH/DoT bypasses it), NETransparentProxyProvider (redirect +
in-OS MITM injection point). Entitlement
com.apple.developer.networking.networkextension is self-service — the most
achievable real-enforcement track.
pf forced egress:pfctl -f -a anchors loaded by a root-owned launchd job the
agent UID cannot flush; rdr-redirect agent-UID egress into the broker. pf is
last-match-wins without quick — verify with pfctl -a … -sr after load or it is
Potemkin. = ADR-0087 phases 5/6; runbook docs/operations/tcb-broker-runbook.md.
Secret detection — NO keyword NLP (hard rule)
Structured key-format anchors + Shannon entropy, never a keyword wordlist classifier.
Shannon entropyH = -Σ p(c)·log₂ p(c) as FALLBACK only — base64 floor ≈ 4.5, hex
≈ 3.0, length ≥ 20, gated on a structured anchor or a known cred path. Never the sole
verdict.
Committed baseline (.pd-secrets-baseline.json, detect-secrets model) — re-scan
surfaces only NEW secrets. Without it the score is noise and trust dies on first run.
Vendor the gitleaks MIT rule corpus (config/gitleaks.toml) as the format library — a
maintained dependency, refreshed on a schedule. The gitleaks "keyword" stanza is a
candidate-narrowing pre-filter ahead of regex, never the verdict — reading it as a
classifier would violate the no-keyword-NLP rule.
Hiding spots: **/.env/.env.*, ~/.aws/credentials, ~/.config/gh/hosts.yml,
~/.netrc, ~/.npmrc, ~/.pip/pip.conf, ~/.docker/config.json (base64 auths.*.auth —
decode bounded fields before matching), ~/.ssh/* PEM, shell history, .mcp.json,
~/.cursor/mcp.json, Claude config. Report path/line/ruleId/last4/entropy — never the
value. Optional live-verify (sts:GetCallerIdentity, GET /user, GET /models)
collapses false positives but SPENDS the secret over the network — default OFF,
operator-consent-gated, routed through the egress meter.
Prior-art reuse stance (license-aware)
REUSE (link-safe): gitleaks rule corpus (MIT); northpolesec/santa rule model +
santactl fileinfo --json trust extractor (Apache-2.0 — google/santa is ARCHIVED as of
Feb 2025, use the fork; an admin-priv bypass was fixed only in Santa 2025.12, pin
versions); osquery tables (Apache-2.0); eslogger (Apple, pre-entitled).
STUDY then reimplement (copyleft — do NOT statically link into PD): LuLu's
NEFilterDataProvider egress-firewall pattern (GPL-3.0); BlockBlock/KnockKnock
persistence taxonomy (GPL-family); trufflehog live-verification idea (AGPL-3.0);
opensnitch rule-DSL (GPL-3.0, the Linux delta). Falco (Apache-2.0) is a rules-DSL design
reference + the Linux/eBPF tool — it does NOT run on macOS.
Existence proofs:redcanaryco/mac-monitor (BSD-3) + Patrick Wardle's tools +
momenbasel/puresnitch (MIT, pf-based) prove an indie can ship a notarized ESF/NE
system extension.
Santa-fronted lockdown (real teeth without holding the ESF entitlement)
Operator installs northpolesec/santa; PD becomes the rule-authoring + posture layer
(StaticRules + an embedded Moroz-style local sync server, Standalone mode = local
approve, no MDM). Real ES_EVENT_TYPE_AUTH_EXEC block-before-exec without PD holding the
Apple-gated ESF entitlement. Rule precedence: CDHash > Binary(sha256) > SigningID > Certificate > TeamID, most-specific-wins; if code-sig validation fails, only file-hash
rules apply. Modes: 1=Monitor, 2=Lockdown, 3=Standalone.
Top risk traps (avoid these specific failures)
"Green = safe" overclaim — the cardinal sin. Green means "cooperative-case sensors
clear," never "sandboxed." The report must say so.
Missing quarantine ≠ safe — it is the dangerous path (curl|bash, npm).
spctl on bare CLI binaries — use codesign/SecCode instead.
ppid attribution — use responsible_audit_token.
ESF AUTH deadline — slow handler is killed; a dead default-deny client soft-bricks
exec. Wrap Santa instead of hand-rolling AUTH_EXEC.
GPL/AGPL contamination — LuLu (GPL-3), trufflehog (AGPL-3) cannot be linked into PD;
gitleaks (MIT), Santa/osquery/Falco (Apache-2), puresnitch (MIT) are safe.
ESF entitlement lead time — Apple-granted, weeks–months, can be declined. Plan the
read-only/eslogger/Santa-front path as the real near-term track.
kill-based "block" is a race — read-only ESF reports an exec already running.
Observe + best-effort kill, never prevention.
Performance/battery — debounce, watch mtimes via FSEvents not full rescans, cache
trust by (path, cdhash), mute platform binaries; or the sniffer becomes the hog it
hunts.
