Operational truth, not vendor hype.

• The operational reality of using Cloud-LLMs for InfraSec and AppSec work.
• Where AI is a brilliant assistant, and where it is artificially stupid or disobedient.
• Field notes from actual usage. American models (Claude from Anthropic). Open-weight Chinese models (DeepSeek).
• Reality vs the marketing hype of their vendors.
• How to reduce AI flaky behavior by separating deterministic checks from the ones that actually need AI thinking.
• How to manage cognitive debt and intent debt.

Who stays accountable at the end of day. (You do.)

Intermediate builders new or limited in AI adoption.

• Why them. Many experienced teams in big software companies are not as AI-embracing, or as AI-assessing, as other teams are. If that is your team, this is for you.

What the advice is sized for.

• Legacy systems.
• Multi-repository projects.
• Stakes higher than a local school project with no end users.

Ron Michael Diokno Khu.

Proud dad to my daughters. Software engineer by profession, and by passion.

Two decades, across domains.

Where the field notes come from.

Company work (LegalMatch).

• Fleet of Claude sessions parked on Jira ticket updates.
• Prod access with sudo for investigation. (Yikes. I know.)
• Merges flow to staging and main branches.
• Mini CI/CD auto-deploys and notifies MS Teams.

Android phone.

• Termux + ssh + tmux + Claude Code.
• Manual steering while away from the laptop.

Community and personal.

• DeepSeek + pi.dev.

Session wrangling (new).

• Now using herdr to organize ssh and AI sessions in one place.
• Flow is now smoother. I do not open a lot of Termux tabs anymore.

How a ticket becomes production.

Every deploy pings MS Teams. herdr organizes the sessions. Phone + Termux + ssh + tmux for mobile steering.

Brilliant and artificially stupid, same session.

The human operator is accountable. End of day.

Newer release does not equal better.

• Field reality: regressions, dumbed-down tool use, over-restrictive refusals.
• Among advanced devs today, Claude is widely considered crippled. DeepSeek is still smarter for a lot of real work, and it is open-weight on top of that.
• Trusting marketing benchmarks leads to fragile pipelines.

Fair disclaimer. Rankings here are a snapshot, not a verdict.

• Can flip. Vendors ship, degrade, and ship again. Re-evaluate per quarter, not per keynote.
• Domain-shaped. Example: Claude builds JavaScript decks and can see its own render. DeepSeek has no vision API and must approximate layout from text alone.
• Takeaway. Match the model to the task. Keep two or three in rotation.

“Is this site secure?”

Months ago I tried the same prompt across several models.

Not a smart vs dumb question. A workflow question. Pick the behavior that fits your workflow and your legal posture.

Safety guardrails that refuse legitimate security work.

• Examples. Refusing to parse a payload capture. Refusing to help on a log that contains PII you own. Refusing to touch an IAM policy because the prompt says "exploit."
• The hidden cost. Hours bled engineering jailbreaks for benign work.

Guardrails Sickness: a vendor-side disease the operator pays for.

You are letting software crawl your infra and tell its server.

Threat-model it like you would any third-party contractor you cannot audit.

Never use an LLM for what a 5-line bash script can do.

The split removes flaky AI from load-bearing paths.

Who is good at what. And who pays.

Accepting AI code you cannot defend.

• Symptom. Your repo passes checks nobody can defend in a code review.
• Compound interest. Six months later it breaks, and nobody on the team knows why that firewall rule exists.

Fix: AI explains. You understand. Then you ship.

What you meant vs what the AI shipped.

• Example 1. Asked for "block external access to admin." Got a rule that blocks external access, except for a hardcoded CIDR the AI invented.
• Example 2. Asked for "least privilege S3 policy." Got a wildcard permission with a comment that says // least privilege.
• Silent security gaps. The code reads right. The intent is wrong.

Fix: verify the enforcement, not the vibes.

Six rules for the operator.

• 1. Least privilege for AI agents. Treat an LLM service account like a junior contractor from a vendor you cannot audit.
• 2. Deterministic to scripts. Probabilistic to the LLM. Never swap.
• 3. Do not paste prod secrets, topology, or customer data into a cloud LLM. If you must send infra context, send a redacted shape.
• 4. For sensitive work, prefer open-weight models you host. You keep the weights. The data does not leave the box.
• 5. Same code-review bar as a dev's. No "it is just a quick fix" exceptions.
• 6. Verify the intent, not the vibes.

The real cost is the operator who stops learning.

• The real cost of AI-first operation is not tokens.
• Use the LLM as a learning amplifier. Ask it to teach you, not just do.
• Then you can audit it. Then you can disagree with it. Then you are actually accountable.

Build yourself up, not just your output.

Local and global. Participate in the discussions.

• Why it matters. No single operator tracks every model regression, every quiet nerf, every new attack vector. The field moves too fast for one person.

My own.

• AWS User Group Cebu
• PizzaPy
• IoT Cebu
• CeGNULUG (Cebu GNU/Linux Users Group)

Participate. Do not just lurk.

Share a field note. Ask the question. File the bug. Challenge the hype. The community gets smarter when operators show up and contribute what they are actually seeing.