v0.9.0 — hardening + product finish

SysSpectogram v0.9.0 — hardening + product finish

v0.9 hardens the single-VPS path without pretending to be kernel EDR, finishes
the operator story (Day-0, golden path, eval, threat model), and cleans up
overloaded README navigation.

Included

Safe artifacts (no pickle on the critical path)

• Train/export write scaler.json and iforest.ssf.npz (+ meta).
• Runtime refuses .joblib; migrate-only legacy read via
  artifacts migrate --delete-legacy.
• Agent IF train/load uses .ssf.npz.
• supply_chain.enforce: true refuses unsigned / unsafe loads.

Remote alert sinks

alerts:
  allow_private_sinks: false
  strict_secrets: true
  sinks:
    - type: file
      path: reports/alerts.jsonl
    # - type: syslog
    #   host: siem.example
    #   port: 6514
    #   transport: tls
    # - type: https
    #   url: https://collector.example/ingest
    #   token_env: SS_ALERT_TOKEN

Fail-open per sink. HTTPS SSRF guards + token_env for secrets. Not a full SIEM.

FIM

• Baseline persists to state/fim-baseline.json (+ .sha256 seal) via serde_json.
• Enabled by default on the lite load profile (longer interval).

Hybrid watchdog

• Phoenix userspace twin + Dead-man remote emit; systemd WatchdogSec.
• Optional DKMS module packaging/kmod/sysspectogram_wd (PID registry;
  watchdog.kernel_protect: false by default).

Narrow auto-isolate

kirk:
  auto_isolate: false
  auto_isolate_host_risk: false
  host_risk_threshold: 0.99

Host-risk path still requires allow_ssh_cidrs.

Docs & eval

• README EN/RU: grouped links, Limitations, ONNX VPS recipe.
• DAY0_VPS.md, GOLDEN_PATH.md,
  EVAL.md, THREAT_MODEL.md,
  SUPPLY_CHAIN_STORY.md, AGENT.md.
• scripts/eval_offline.py, scripts/demo_golden_path.sh.

Migrate existing models

python -m sysspectogram artifacts migrate --model artifacts/live --delete-legacy
python -m sysspectogram supply-chain manifest --root artifacts/live
# re-sign if you enforce

Verify

pytest -q
cd agent && cargo check
python scripts/eval_offline.py --dataset dataset/ --methods zscore --out reports/eval

Honest ceiling

Same-host root can stop units and wipe local disks. Remote sinks + signatures
make that louder and harder to hide — not impossible. VMI remains v1.0.

SysSpectogram v0.8.0 — signed supply chain

SysSpectogram v0.8.0 — signed supply chain

v0.8 hardens the path from a release or model builder to a running VPS:
signed artifacts, deterministic manifests, SPDX SBOM, distro package recipes,
and a safe model-loading policy.

Included

• Minisign signatures for release binaries, Python distributions, checksums,
  SBOM, and profile packs.
• Deterministic artifacts.manifest.json with SHA-256 and size for every model
  file.
• Optional supply_chain.enforce policy propagated through guard, monitor,
  offline analysis, web, and Telegram scoring.
• cnn.pt loading with weights_only=True only; no unsafe arbitrary-object
  fallback.
• SPDX 2.3 SBOM generator for installed Python and Cargo dependencies.
• Dependabot configuration for pip, Cargo, and GitHub Actions.
• Buildable AUR and Debian packaging recipes with hardened systemd units.
• Release verification helper that never executes downloaded assets.

Compatibility and rollout

Existing v0.7.x configs continue to work with enforcement disabled:

supply_chain:
  enforce: false

After installing the release public key and signing the model directory:

supply_chain:
  enforce: true
  public_key: /etc/sysspectogram/sysspectogram.minisign.pub
  manifest_name: artifacts.manifest.json
  signature_name: artifacts.manifest.json.minisig

Use python -m sysspectogram configure to set this interactively. It refuses
to enable enforcement without a public key.

Verify release assets

minisign -Vm sysspectogram-agent-x86_64-linux-musl \
  -p SysSpectogram.minisign.pub \
  -x sysspectogram-agent-x86_64-linux-musl.minisig

scripts/verify-release.sh SysSpectogram.minisign.pub \
  sysspectogram-agent-x86_64-linux-musl \
  sysspectogram-agent-x86_64-linux-musl.sha256

Verify model artifacts

python -m sysspectogram supply-chain verify artifacts/real_v3 \
  --enforce --public-key /etc/sysspectogram/sysspectogram.minisign.pub

--insecure remains available for explicit local profile installation only;
it is not a production trust mode.

Build and package

Builder:

python -m build
python scripts/generate-sbom.py --out dist/sbom.spdx.json
cd agent && cargo build --release --locked --no-default-features

Arch:

cd packaging/aur
makepkg -si

Debian/Ubuntu:

dpkg-buildpackage -us -uc -b -d

The packaged agent defaults to userspace mode. eBPF requires a compatible
kernel and the documented capabilities; it is not silently assumed.

Upgrade

git fetch --tags
git checkout v0.8.0
python -m venv .venv
source .venv/bin/activate
pip install -e '.[onnx]'
cd agent && cargo build --release --locked
cd ..
python -m sysspectogram configure

Keep response.mode: observe and kirk.auto_isolate: false during the first
upgrade, verify the public key and model signature, then restart the services.

Security notes

iforest.joblib and scaler.joblib remain legacy pickle-compatible files.
They are accepted only after manifest verification when enforcement is enabled.
Prefer ONNX inference on small VPS hosts and generate model artifacts on a
trusted builder. See SUPPLY_CHAIN.md and SECURITY.md.

Known ceiling

kirk.trust: best-effort remains honest on hosts without IMA plus measured
boot evidence. Live VMI is still conditional v1.0 functionality for self-hosted
KVM and is not part of this release.

v0.7.1 — module_hide FP + Ignore mute

SysSpectogram v0.7.1 — module_hide FP + Ignore mute

Fixes

• agent_kirk_module_hide false positive: /sys/module lists kernel builtins (acpi, 8250, …) that never appear in /proc/modules. Cross-view now only compares loadable modules (those with /sys/module/*/initstate), so builtin noise no longer floods Telegram.
• Ignore button for kirk/host alerts: clicking Ignore on an alert without a process identity used to ack "ignored" without storing anything. It now mutes by rule_id (persisted in state/process_labels.json), and the guard skips muted rules.

Upgrade

git pull
pip install -e '.[onnx]'   # or your usual install
cd agent && cargo build --release
# restart agent + guard

If spam already started, either Ignore once more (now persists) or:

python -c "
from sysspectogram.process_labels import ProcessLabelStore
from pathlib import Path
s = ProcessLabelStore(Path('state/process_labels.json'))
s.mute_rule('agent_kirk_module_hide', note='builtin-fp')
print('muted')
"

v0.7.0 — Deeper ops on the VPS

SysSpectogram v0.7.0 — Deeper ops on the VPS

Release after v0.6 agent harden: Day-0 configure, quieter false positives via
ProcessLabelRules, richer digests/audit, cheaper flow, optional IF refit —
without forcing CNN fine-tune on a 1 GB box.

Why it matters

Before (v0.6)	Now (v0.7)
Day-0 = edit YAML + thin setup for .env	configure TUI + host_probe recommendations; overrides need confirm
As normal / As anomaly = bias + vague retrain	ProcessLabelRules (exact / prefix / glob / …) + /labels; role FP seeds
Full CNN retrain expected somehow on VPS	Rules first; optional IF-only refit; CNN stays builder-only
Flow = /proc/net/tcp only	netview (ss) + /proc fallback on full profile
Destructive TG actions hard to audit	reports/response_audit.jsonl
Kirk trust only in console/TG	Web kirk badge on live dashboard
Root = ProcWatcher poll	+ eBPF setuid→0 assist (agent_ebpf_setuid_root); ProcWatcher remains fallback
Cold install tribal knowledge	COLD_INSTALL.md checklist

Honest limits: labels do not replace a good host model; IF refit is opt-in and still not a full detector rewrite; eBPF setuid needs root/CAP_BPF and quiet hosts (desktop flood remains a reason to prefer userspace). Trust stays best-effort without IMA+SB/TPM.

What's new

Configure (killer Day-0)

• python -m sysspectogram configure — Rich terminal UI (not a browser)
• host_probe: RAM / vCPU / container / desktop / onnx|torch → recommend lite/full, agent, flow, runtime.prefer, feedback knobs
• Changing away from recommendation shows a warning + y/N
• --accept-recommended --role ssh for scripted / cold VPS
• Writes deep-merged configs/default.yaml, .env, state/host_probe.json
• Optional role FP seed into state/process_labels.json

Docs: CONFIGURE.md · COLD_INSTALL.md

ProcessLabelRules + feedback loop

• Store: state/process_labels.json (v2). Legacy operator_feedback.json migrates.
• Match: exact, comm_prefix, comm_regex, path_glob, path_contains, cmdline_contains
• Priority: anomaly > ignore > baseline
• TG: As normal / As anomaly / + similar, /labels, /label-del
• CLI: labels list|seed|del
• Role seeds: ssh / nginx / docker / panel / python / wireguard (role_fp)
• Global threshold bias via FeedbackLearner kept
• Optional: feedback.if_refit: true → debounced / CLI feedback retrain-if (CNN/ONNX untouched)
• Builder full retrain: feedback retrain (Torch) unchanged

Docs: FEEDBACK.md · TELEGRAM.md

Flow / netview (Track B)

flow:
  enabled: true          # full preset
  backend: netview       # proc | netview | both
  window_sec: 30
  syn_threshold: 80
  unique_port_threshold: 40

Lite keeps flow off / proc by default.

Response UX (Track C)

• Append-only reports/response_audit.jsonl (no secrets in payload)
• Lockdown / isolate still go through unlock + confirm tokens
• Live web: kirk trust badge + isolated flag (web static)

eBPF setuid assist (Track E)

• Probes: setuid / setreuid / setresuid toward uid 0 from non-root
• Alert id: agent_ebpf_setuid_root
• Lite / no-eBPF: ProcWatcher root_watch still authoritative

Docs: EBPF_SETUP.md · ROOT_WATCH.md

Digests / ops polish (Track A)

• /digest includes alerts_today, response_mode, kirk trust, feedback bias/counts, label count, last IF refit
• Version bump: package + agent 0.7.0

Quick start

pip install -e ".[onnx]"   # or .[dev] / .[ml] on builder
cd agent && cargo build --release && cd ..

python -m sysspectogram configure --accept-recommended --role ssh
# or interactive:
# python -m sysspectogram configure

python -m sysspectogram guard --model artifacts/real_v3 --telegram --dry-run
# console prints UNLOCK CODE → TG: /unlock NNNNNN
# /labels  /digest  As normal / + similar

Optional IF refit (after enough labeled windows):

python -m sysspectogram feedback retrain-if --model artifacts/real_v3

Config knobs (new / important)

load_profile: lite   # or full — configure / probe picks this
runtime:
  prefer: notorch    # torch only if probe says so
agent:
  enabled: true
  auto_start: true
  mode: userspace    # ebpf on quiet VPS with root
  root_watch: true
flow:
  enabled: false
  backend: proc      # netview on full
feedback:
  if_refit: false
  widen_rules: comm_prefix   # false | comm_prefix | path_glob | both
response:
  mode: observe
telegram:
  require_console_unlock: true

Docs

• CONFIGURE.md · COLD_INSTALL.md · FEEDBACK.md
• ROADMAP_QUALITY.md — v0.7 shipping; next v0.8 supply chain
• RELEASE_v0.6.0.md — prior agent harden release
• README EN/RU updated for configure Day-0

Upgrade notes

1. pip install -e '.[onnx]' (or reinstall) — __version__ → 0.7.0.
2. Rebuild agent (cargo build --release) for setuid probes + prior root_watch.
3. Run configure once (or merge YAML knobs above by hand).
4. Old state/operator_feedback.json → labels migrate on first guard/labels load.
5. Prefer agent.mode: userspace on desktop; ebpf on quiet VPS as root.
6. Keep --dry-run / response.mode: observe until unlock + labels are trusted.
7. Do not expect CNN fine-tune on 1 GB VPS — use labels + optional IF; full retrain on a builder PC.

Not in 0.7

• CNN fine-tune on VPS / Rust CNN training
• minisign / cosign / SBOM (v0.8)
• Live VMI (v1.0)

v0.6.0 — Harden the agent path

SysSpectogram v0.6.0 — Harden the agent path

Release after v0.5 Adopt/Kirk: make the same-UID / lite VPS path honest and usable —
auth, self-protect, unexpected root in Rust, Role Lab packs, operator feedback.

Why it matters

Before (v0.5)	Now (v0.6)
Critical kirk alerts forgeable on same UDS	HMAC + PEERCRED fail-closed + PID allowlist + exe seal
Agent death can go quiet	Phoenix watchdog, CLEAN_SHUTDOWN, Dead-man, TG bypass
Root process watch = Python /proc (heavy on lite)	Rust ProcWatcher same walk → agent_unexpected_root + TG Kill
Unexpected SSH soft	Learn window → Kick tty / Ban IP
Packs unsigned locally	HMAC .sig + validate-pack
Retrain feedback vague	TG Ignore / As normal / As anomaly → learner (+ retrain CLI)
Role baselines hard without VMs	Role Lab collect/train/pack on a PC

Honest limits: cannot revoke uid=0 without Kill; root can still replace a poorly installed binary — install under root-owned /usr/local/sbin. Cloud IMA/SB often absent → trust stays best-effort.

What's new

Agent auth and seal

• HMAC-SHA256 for critical rules (state/agent_hmac.secret)
• Same-UID PEERCRED fail-closed; refresh allowlist from agent + watchdog PIDs
• Binary seal (state/agent_binary.seal.json); mismatch → loud alert

Docs: AGENT_PROTECT.md · SECURITY.md

Unexpected root (lite-safe)

• agent.root_watch: true (default) — detection in agent, no second /proc walk
• Python root_watch only if agent.enabled: false
• Rule: agent_unexpected_root (HMAC required)

Docs: ROOT_WATCH.md

Sessions, Role Lab, feedback, OSINT

• ROLE_LAB.md · FEEDBACK.md · OSINT.md · DAY0_VPS.md

CI / packaging

• GitHub Actions pytest + agent checks; release-assets workflow
• packaging/sysspectogram-agent.service

Quick start

pip install -e ".[dev]"
cd agent && cargo build --release && cd ..
python -m sysspectogram kirk trust
python -m sysspectogram guard --telegram --dry-run
# console unlock code → TG /unlock NNNNNN

Agent (root, optional eBPF):

sudo install -m 0755 -o root -g root \
  agent/target/release/sysspectogram-agent /usr/local/sbin/sysspectogram-agent
sudo /usr/local/sbin/sysspectogram-agent --mode ebpf --phoenix \
  --socket /run/sysspectogram/agent.sock \
  --hmac-secret /opt/sysspectogram/state/agent_hmac.secret

Config knobs (new / important)

agent:
  enabled: true
  require_hmac: true
  require_same_uid: true
  root_watch: true
  root_learn_sec: 300
root_watch:
  enabled: false   # Python fallback only without agent
sessions:
  enabled: true
  learn_sec: 120

Docs

• ROADMAP_QUALITY.md — forward plan (v0.7+)
• AGENT.md · KIRK.md · TRUST.md
• RELEASE_v0.5.0.md — prior Adopt/Kirk release

Upgrade notes

1. Rebuild agent (cargo build --release) — root watch + HMAC live in the binary.
2. Ensure state/agent_hmac.secret shared by guard and agent.
3. Prefer systemd unit under /usr/local/sbin (AGENT_PROTECT.md).
4. Keep kirk.auto_isolate: false until allow_ssh_cidrs is set.

SysSpectogram v0.5.0 — VPS Adopt + Kirk trust

SysSpectogram v0.5.0 — VPS Adopt + Kirk trust

Biggest release since v0.4: turn a hybrid ML host IDS into practical VPS defense you can run on a cheap box without PyTorch, with honest kernel-integrity labels and safer response actions.

Why it matters

Before (v0.4)	Now (v0.5)
Torch often blocked 1GB VPS	runtime: notorch / onnx — day-0 without CNN, or light ONNX
Retrain pain on the server	Train bridge: collect on VPS → train on PC → artifacts push back
“Rootkit” ≈ heuristics	Kirk: module eBPF, cross-view hide, kallsyms presence baseline
Unclear trust	kirk.trust: best-effort | measured (IMA + Secure Boot/TPM) — no hype
IP bans only	kirk_isolate / kirk_release with TTL; TG /kirk_release (unlock-gated)
Flat auto-ban	response.mode: observe → shield → aggressive

Honest ceiling: many Arch/cloud hosts have no IMA in-kernel and Secure Boot off → trust stays best-effort. That is correct. Live VMI (hypervisor) is deferred to v1.0.

What's new

VPS Adopt

• runtime: notorch | onnx | torch_ml (SYSSPECTOGRAM_RUNTIME)
• ONNX export/infer; train bridge (data bundle / artifacts push with checksum verify)
• Setup/bootstrap helpers

Kirk (in-guest integrity)

• eBPF execve/openat/module; cross-view hide; --kirk-seal presence baseline
• python -m sysspectogram kirk trust (+ reason codes)
• Docs: KIRK.md · VMI.md (VMI → v1.0)

Response & safety

• nft ss_kirk isolate + TTL; refuse empty allow_ssh_cidrs
• auto_isolate allowlisted CRITICAL only (default off)
• Hardened push paths; no false measured on unreadable IMA

Quick start

pip install -e ".[dev]"
cd agent && cargo build --release && cd ..
python -m sysspectogram kirk trust
python -m sysspectogram guard --telegram --dry-run

# ONNX
SYSSPECTOGRAM_RUNTIME=onnx python -m sysspectogram guard \
  --model artifacts/real_v3 --telegram --dry-run

# eBPF agent (root)
sudo -E ./agent/target/release/sysspectogram-agent \
  --mode ebpf --socket "$XDG_RUNTIME_DIR/sysspectogram-agent.sock"

Profile pack

sha256: a8402315fec15ed9dd602f78dcbb8b59c680768d7f904e46175dd22d3abb20b4

python -m sysspectogram profiles pull \
  --url https://github.com/lifeqsoll/SysSpectogram/releases/download/v0.5.0/profile-generic-linux-v1.tar.gz \
  --out /tmp/p.tar.gz \
  --sha256 a8402315fec15ed9dd602f78dcbb8b59c680768d7f904e46175dd22d3abb20b4

python -m sysspectogram profiles install /tmp/p.tar.gz \
  --dest artifacts/profiles/generic-linux \
  --sha256 a8402315fec15ed9dd602f78dcbb8b59c680768d7f904e46175dd22d3abb20b4

Docs

• ROADMAP_QUALITY.md
• TRAIN_BRIDGE.md
• KIRK.md
• VMI.md
• CONFIG.md

Notes

• Default runtime is notorch (host CNN off until onnx/torch_ml + --model).
• Do not enable kirk.auto_isolate without allow_ssh_cidrs.
• Seal kallsyms as the same user that runs the agent.

v0.4.0

SysSpectogram v0.4.0

Linux hybrid host IDS: CNN + Isolation Forest, perimeter/Telegram SOAR-lite, live web / Mini App, and a Rust integrity agent.

What's new

• Fuse risk — host ML + agent IF (ensemble.host_weight / agent_weight)
• Profile packs — shareable tar.gz baselines (profiles pack / pull / install)
• Load budgets — lite (small VPS) / full (large VDS); env SYSSPECTOGRAM_LOAD_PROFILE
• FIM — optional sha256 path watches (full)
• Flow lite — /proc/net/tcp heuristic (full)
• Console unlock — TG /unlock + Mini App gate (stolen .env alone cannot kill/ban)
• eBPF — optional execve / openat via clang BPF + Aya loader (attach needs root; on Arch setcap alone is not enough)

Assets

File	Purpose
profile-generic-linux-v1.tar.gz	Baseline host + agent-IF pack
profile-generic-linux-v1.tar.gz.sha256	Integrity check (required by profiles pull/install)

python -m sysspectogram profiles pull \
  --url https://github.com/lifeqsoll/SysSpectogram/releases/download/v0.4.0/profile-generic-linux-v1.tar.gz \
  --out /tmp/p.tar.gz \
  --sha256 a8402315fec15ed9dd602f78dcbb8b59c680768d7f904e46175dd22d3abb20b4

python -m sysspectogram profiles install /tmp/p.tar.gz \
  --dest artifacts/profiles/generic-linux \
  --sha256 a8402315fec15ed9dd602f78dcbb8b59c680768d7f904e46175dd22d3abb20b4

Quick start

pip install -e ".[dev]"
cd agent && cargo build --release && cd ..
python -m sysspectogram guard --model artifacts/real_v3 --telegram --dry-run
# eBPF agent:
# sudo -E ./agent/target/release/sysspectogram-agent --mode ebpf --socket "$XDG_RUNTIME_DIR/sysspectogram-agent.sock"

Docs

• PROFILES.md
• EBPF_SETUP.md
• AGENT.md
• ROADMAP_V3.md

SysSpectogram v0.2.0 - hybrid host ML + perimeter + Telegram

Tag: v0.2.0
Title: SysSpectogram v0.2.0 — hybrid host ML + perimeter + Telegram

--- paste below into GitHub Release description ---

SysSpectogram v0.2.0

First public release of a Linux hybrid IDS: host behavioral ML (CNN + Isolation Forest) plus perimeter/egress guard and a Telegram SOAR-lite control plane.

Repo: https://github.com/lifeqsoll/SysSpectogram

Highlights

• Host pipeline: collect → build-dataset → train → monitor / analyze
• Perimeter: SSH auth fail/brute, inbound scan heuristics, egress denylist, suspicious DNS
• guard: perimeter + host ML + Telegram slash/inline (confirm before ban/kill/lockdown)
• Dual alert panel (metrics heatmap + top-PID CPU)
• OSINT recon (PTR/enrich/DNS/CT); nmap only via lab allowlist
• Local load sims + intrusion lab scripts (localhost only)
• Safer defaults: recon/nmap off, TG simulate/collect off, nft unban by handle, temporal val split, artifact checksums, webhook HMAC

Install

git clone https://github.com/lifeqsoll/SysSpectogram.git
cd SysSpectogram
python -m venv .venv && source .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -e ".[dev]"
cp .env.example .env   # optional Telegram

Quick start

# train your own artifacts (not shipped)
python -m sysspectogram collect --out data/normal.csv --duration 1800
# …sim + anomaly collect…
python -m sysspectogram build-dataset --normal data/normal.csv --anomaly data/anomaly.csv --out dataset/real
python -m sysspectogram train --dataset dataset/real --out artifacts/real

# live (dry-run safe)
python -m sysspectogram guard --model artifacts/real --telegram --dry-run

Docs

• README · RU
• Recipes · Telegram · Config

Safety

• Linux only. No Win/Mac.
• Destructive TG actions require confirm; ban/kill need real privileges (omit --dry-run).
• Do not aim sims/nmap at third-party networks.
• Model quality = your labeled data; no pretrained weights in the repo.

Known limitations

• Live nft without dry-run needs suitable privileges/systemd caps
• Retrain to get checksums.sha256 on artifacts
• Not antivirus / full NIDS / SIEM