Host Security Baseline Audit (Linux + Windows)
A defensible posture report instead of a scanner's opinion.
$49
Works with the AI Agents you already use
Host Security Baseline Audit (Linux + Windows)
Example session with this skill installed
Audit our Ubuntu web server web-01 against the baseline before the customer's security questionnaire is due. I have sudo. Tell me what to fix first and what evidence closes each item.
- Read your context and instructions
- Compiled the host security baseline
Summary
FAIL on web-01: 13 pass, 5 fail, 1 unknown out of 19 applicable Linux controls (16 Windows controls not applicable). Two high-severity SSH and audit items and one exposed service lead the list. The unknown (ptrace scope) needs one value before the report is complete.
$ sudo bash scripts/collect_linux.sh --out obs-web-01.json
wrote obs-web-01.json
$ python3 scripts/evaluate.py --observations obs-web-01.json
FAIL platform=linux host=web-01.example.internal collected=2026-10-29T12:00:00Z
pass=13 fail=5 unknown=1 notApplicable=16 (unknown never counts as pass)
Fix first (severity order)
| Rule | Sev | Finding | Fix | Closes with |
|---|---|---|---|---|
| HB-002 | high | 1 risky service listening beyond loopback: Redis on 0.0.0.0:6379 | bind Redis to 127.0.0.1 (bind 127.0.0.1 in redis.conf), restart; firewall the port | ss -ltn shows 6379 on loopback only |
| HB-201 | high | PermitRootLogin prohibit-password (not no) | named admin accounts with sudo exist; set PermitRootLogin no, reload sshd | sshd -T reports permitrootlogin no; negative direct-root login test |
| HB-211 | high | auditd inactive | apt install auditd audispd-plugins, enable with an approved ruleset, forward logs | systemctl is-active auditd and a recorded test event |
| HB-204 | medium | MaxAuthTries 6 | set 4 after confirming automation does not offer more than 3 keys | sshd -T reports maxauthtries 4 |
| HB-216 | medium | root filesystem on /dev/vda1, not encrypted | record the cloud provider's volume-encryption guarantee as a compensating control, or rebuild with LUKS | documented exception with approver, or lsblk shows a crypt layer |
Unknown to resolve
| Rule | Reason | Action |
|---|---|---|
| HB-210 | kernel.yama.ptrace_scope not present (Yama LSM not available on this kernel) | record as not_applicable with reason "Yama not built into kernel 5.15-cloud" and the approver, or switch to a kernel with Yama |
Passing (13)
Disk headroom 42%, time synchronized, password auth off, empty passwords off, firewall active (nftables), IPv4 forwarding off, /etc/shadow 640, full ASLR, SUID dumps off, AppArmor enforcing, unattended security updates on, no NOPASSWD: ALL sudoers, no unexpected world-writable files.
Evidence pair for the questionnaire
obs-web-01.json (the observations, as collected) and evaluation-web-01.json (--json output). Re-run after the fixes; attach the second pair showing 0 fail, 0 unknown, and the one documented exception.
Connects securely to your tools. The creator never sees your data.
About this skill
A defensible posture report instead of a scanner's opinion. Thirty-five controls that matter on almost every host, each with the impact, the remediation, the closure evidence and NIST CSF 2.0 mappings, evaluated deterministically from observations your agent collects with read-only scripts.
Controls: common (disk headroom, risky services listening beyond loopback, time sync); Windows (UAC, firewall profiles, Defender real-time and signature age, RDP NLA, SMBv1 off, SMB signing, PowerShell script-block logging, LSA protection, Secure Boot, Windows Update service, BitLocker on the OS volume, guest account, LAPS, logon auditing, local admin count); Linux (SSH root login, password auth, empty passwords, MaxAuthTries, host firewall, IPv4 forwarding, /etc/shadow mode, full ASLR, SUID core dumps, ptrace scope, auditd, SELinux/AppArmor enforcing, unattended security updates, NOPASSWD:ALL sudoers, world-writable files, encrypted root).
Scripts:
collect_linux.sh(bash, read-only, root for complete results) andcollect_windows.ps1(PowerShell, read-only, elevated for complete results): one JSON document each, withunknownplus a reason when a value could not be read, never a guess.evaluate.py(Python stdlib): pass / fail / unknown / notApplicable per rule, severity-ordered report with fix and closure lines,--jsonfor diffing runs. Exit 0 clean, 1 findings, 2 could not evaluate (which is not a worse 1: it means the instrument did not run).
Reference: the rulepack (HB-001...HB-216), a generated control catalog, and sample observation files for both platforms.
No network access. A baseline, not a CIS benchmark; it tells you what to fix first and how to prove it is fixed.
35-control baseline only. Linux collector optimized for Debian/Ubuntu and RHEL. Requires root/admin elevation for full visibility. Measures configuration, not real-time behavior.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 19 days ago
Requires