Wiki › Engineering with agents

Baseline Security Controls

Cheap, unglamorous controls that close the paths most attacks use: unique passwords and passkeys, patching, supply-chain hygiene, segmentation, verification, and tested backups.

Baseline security controls are the unglamorous, cheap measures that close the paths most attacks actually use: reused passwords, unpatched software, flat networks, untested backups and people in a hurry. The Markets & Mayhem guide "A Practical Guide to Improving Your Security in 2026" orders them the way an attacker meets them and writes them for both beginners and network operators. Its thesis is that attackers look for the cheapest path, so defence should remove the cheap paths and make the remaining ones noisy.¹

Controls in order

• Credentials. Use a password manager and unique, long passwords (NIST: at least 15 characters when the password is used alone; no composition rules). Change passwords only when there is evidence of compromise; NIST and the UK NCSC now reject scheduled rotation. Watch for breaches with Have I Been Pwned.
• MFA. Order of strength: passkeys or hardware keys first (phishing-resistant), then TOTP, then push with number matching, then SMS, then email codes. MFA is usually bypassed through recovery and enrolment paths: push fatigue, help-desk impersonation, SIM swaps and reverse proxies. Remove weak fallbacks and register two keys.
• Patching. Turn on automatic updates for personal devices. Organisations should use a pilot ring and then a broad ring. Treat a device that no longer gets updates as end of life.
• Software supply chain. After Shai-Hulud-style npm worms:

- restrict install scripts (ignore-scripts, allowScripts);

- install from lockfiles;

- wait 3–7 days before adopting a release (minimumReleaseAge);

- check provenance;

- publish with short-lived OIDC tokens;

- keep secrets out of build environments;

- run Node LTS one major version behind.

• Network. Disable UPnP, WPS and remote router administration. Default-deny inbound, with egress filtering for advanced users. Add DNS filtering (NextDNS, Quad9, Pi-hole). Put guests and IoT devices in separate segments or VLANs, and test that they really cannot reach the main computers.
• Devices and data. Turn on full-disk encryption and escrow the recovery key. Prune app permissions and OAuth grants. Encrypt sensitive files before uploading them to the cloud (Cryptomator, VeraCrypt). Keep keys in a secret manager, never in notes.
• People. Against deepfakes (for example, the roughly $25M Arup case), use a shared trust phrase that is never written down, call back on a number you already know, require dual approval for payments, and verify bank-detail changes separately.
• Backups and drills. Follow 3-2-1 plus one offline or immutable copy, with separate credentials, encryption, and a quarterly restore test that measures recovery time. Run tabletop drills for a lost laptop, ransomware and a fraudulent payment request.
• AI for self-audit. Treat the agent as an untrusted analyst: read-only copies of data, a disposable VM, no credentials, restricted egress and capped spend. Beware prompt injection and the "lethal trifecta" (private data, untrusted content and external communication).¹

The guide also has a 30-day plan for beginners:

1. Week one: password manager, MFA, automatic updates.

2. Week two: disk encryption, app review, breach alerts.

3. Week three: backups, network segmentation, DNS filtering.

4. Week four: a verification protocol and a drill.¹

Assessment

The guide compiles mainstream guidance from NIST, CISA, NCSC, FIDO and OWASP with sources. It is not original research, and its author is a markets commentary account. It is useful as a checklist, and it is the "raise the floor" half of the response to the AI Vulnerability Tsunami: automated attackers make unpatched, internet-reachable and reused-credential systems the first casualties. The AI self-audit advice mirrors the read-only, isolated-environment design in Agentic Security Remediation and the risks discussed in Production LLM Reliability.