benarent

The @goteleport getting in the sandbox game. Lots of great features baked in. beams.run

It’s ironic that “Delve” was a common LLM filler word circa 2023

▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓░ 99.9%

American Healthcare: A: Immediate Telehealth care via an App B: We only accept appointments via a phone call, 9 - 5

Claude Code was finally able to find why my Windows 11 box was misbehaving. Lot's of great power-shell debugging.

@ciaran_lee We all know how this ends

Just had an urge to tell a cloth to wipe up my spilled coffee. Too much Claude Code and Wispr Flow and too little sleep will do strange things to a man.

GN SF

Uber going back to its roots.

@stuartchaney plot twist; each PR is +20k

@benarent give me a stanky +20,000 PR at 3:59pm Fri

after almost 20 years of writing code every day, I just hit my personal PR record 20 meaningful updates merged in under 8 hours something big on the way..

I plan a follow up blog post; I've a full report of issues and resolutions available trust.goteleport.com/resources?s=ac…

We just wrapped up a red team exercise with @Persistent_Psi to try to bypass @goteleport Session Summaries. goteleport.com/platform/ident… These are some things I learnt along the way: - Modern LLMs are much better at 'classic' attacks, prompt injections - Since LLMS understand a range of languages can decode base64; you can't just whitelist/blacklist commands - Constrain your LLM output formats and validate deterministically. Expect a number? Check it's a number. - Most AI security issues come from YOLO. throwing everything into the LLM with no constraints.

Oakland Owl

In a couple of weeks I'm hosting an agent forward Identity Security Dinner. Still a few spots left luma.com/t9zoaekx

@ebloch I recommend using the codex MCP in Claude.. then using /codex to pass on harder problems or get a plan reviewed

Finally getting my "health" brain setup. Running Claude Code (in terminal) w/ Opus 4.6 High side by side with Codex 5.4 High (codex MacOS). First time really playing with codex MacOS. Right off the bat really liking the folder/task UI - easier to work across folders and tasks, pun intended. And the app itself feels very smooth compared to Claude mac os, which is too slow for me to use hence I'm terminal boy with Claude Code (and mostly love it). More notes to follow soon on tooling and what i'm actually doing.

@spikebrehm Relatable

finally got sick of painfully slow Next.js hot reload times, and issues with turbopack that forced us back to webpack we kicked off a Cursor cloud agent in GRIND MODE to migrate us to TanStack Start and 14 hours later it's still grinding…

Looking forward to @BSidesSF and the chaos + fun of @OneRSAC.

Great post by Justin Poehnelt on rewriting your CLI for AI Agents. justin.poehnelt.com/posts/rewrite-…

colleague sent me this.

it's ironic that SightGlass doesn't / hasn't provided Wifi since opening.. maybe this is just a device to ear hotspot