Chevy Phillip

AppSec engineer building secure AI agents in public. Built + more. Every number is real.

[Chevy Phillip @chevonphillip]

Since I started running Grok Bot with Cursor Cloud Agents, I haven't opened Codex or Claude Desktop in days. Feels weird, but I'm more productive and more in control.

So what about the Claude usage I'm already paying for? @thecsguy's Claude Remote template puts it to work. Grok Bot hands tasks to Claude Code on my Mac and brings the results back.

Anyone else stopped opening their coding apps?

[nick @thecsguy]

if you have claude code on a mac, you need this grok bot.

text it a task from anywhere, it runs it in a claude code session and pings you with the results. reply and it goes to the same session.

grab the template 👇
https://t.co/DTpPx14Adv

[Refuseless @refuselessai]

Since you're loving DeepSeek... WE SLASHED PRICES EVEN MORE

$0.03 in / $0.10 out / $0.003 cache

[Chevy Phillip @chevonphillip]

@cltgrowth Fair point. Asking the people who'd actually buy beats guessing from a small sample. Appreciate it.

[Chevy Phillip @chevonphillip]

Agent Wins update. Launched Oct 13 with 34 sourced wins across 9 agents. 45 visitors Oct 13-16. 0 sales so far. Not pretty, but real, and building in public means posting the zeros too. If you were me, what would you change first? https://t.co/2En4TEPXsI

[Chevy Phillip @chevonphillip]

@yeemio Back to back. At this point the patch notes are my bedtime reading.

[Chevy Phillip @chevonphillip]

Hey hackers…this is a message to you. Us Security folks have families to feed and sleep to catch up on. Please turn it down a notch. Between PHP and now NodeJS my butt hurts; I’ve been sitting all day.

[XChat @chat]

Heard you want to chat with Bot

[Chevy Phillip @chevonphillip]

Approval gates only work if the human still reads them. Ask for approval on everything and people click yes on autopilot, which might be worse than no gate. My take: gate the irreversible stuff (send, pay, delete, share) and let the rest flow. Where do you draw the line?

[Chevy Phillip @chevonphillip]

@NetworkChuck The hardest thing for us is tracking down the owners of our PHP applications. Also discovering that some of those app still run on PHP 5…. Yeah it’s been a day.

[Chevy Phillip @chevonphillip]

Try Grok @bot

[nick @thecsguy]

Try Grok @bot

[Chevy Phillip @chevonphillip]

Now this what I’m impressed about for today.

[eric zakariasson @ericzakariasson]

some pages ive been finding useful

- launch burn-down lists
- design docs i iterate on with @bot
- morning briefs
- what did i get done this week

[Chevy Phillip @chevonphillip]

@valerioamirani Good call. Find the drop-off before touching the product, so I know whether it's the top of the page or the offer that's losing people.

[Chevy Phillip @chevonphillip]

@reggiescold Fair point, that's too small a sample to blame the offer yet. Telling one win as a story with the agent front and center is a cheap test. Good luck with the startup planets.

[Chevy Phillip @chevonphillip]

@shanto12 Exactly. Expiry turns a leaked key into a short window instead of an open door, and you don't have to notice the leak in time for it to stop working.

[Chevy Phillip @chevonphillip]

Underrated agent control: network isolation. If a prompt injection lands, what the agent can reach decides how bad it gets. Private network by default, explicit egress, nothing exposed publicly. Guardrails are nice, but I'd rather have a wall. What can your agent's box talk to?

[Chevy Phillip @chevonphillip]

@PeilvDog_CN Parallel for a bit. The old job keeps running read-only so I can diff its output against the new one before trusting the writes.

[Chevy Phillip @chevonphillip]

Agent email, transactional email, calls and SMS, coding-agent seats, RSS feeds.

Normally that's 5 vendors, 5 bills, 5 dashboards.

For me it all runs through one Grok Bot plan. $200/mo. 🧵

[Chevy Phillip @chevonphillip]

@PeilvDog_CN Different model. The judge sits above the workers on its own model, so it isn't grading its own family's habits. Like the router rule, the writer never gets to close its own task.

[Chevy Phillip @chevonphillip]

Each of my bots has one job: PM, TPM, engineer, growth, security lab, X manager. My opinion: one do-everything agent is harder to trust and harder to debug than a team with narrow roles. Generalist agent or specialist team, which are you running?

[Chevy Phillip @chevonphillip]

My Show HN got flagged because the account was brand new. Fair rule honestly, it keeps spam out. Lesson: build the account and the reputation before you need them, not on launch day. Anyone else get burned launching from a fresh account?

[Chevy Phillip @chevonphillip]

@shanto12 Agreed. Egress decides where the agent can go, and scoped keys decide what it can do once it gets there. You want both, and short-lived creds so a leaked one dies on its own.

[Chevy Phillip @chevonphillip]

@k_firsov @TownPerson2 Rough.

[Chevy Phillip @chevonphillip]

@gagu_sh That's the failure mode you want. A loud connection error you can alert on beats a quiet request to something internal. And one proxy gives you a single place to log every egress attempt.

[Chevy Phillip @chevonphillip]

@rez0__ @LiveOverflow Woh is that a new trick? I gotta try that.