Superagent
624 posts
Superagent
@superagent_ai
Red teaming for AI agents (YC W24)
San Francisco Katılım Mayıs 2023
21 Takip Edilen2.5K Takipçiler
Superagent retweetledi
every harness out there is built for gpt and claude. nobody builds for @grok. so i did.
grok-code-fast and grok-4.2 match frontier. not almost. match.
the gap people keep talking about doesn't exist. the tooling gap does.
homanp@pelaseyed
Connect Grok CLI to telegram and use at as a personal assistant.
English
Protect your agents from phishing, prompt injections and other attacks hidden in emails.
Excited to partner with @resend on this!
homanp@pelaseyed
just shipped: Brin + @resend. scores inbound emails for phishing, injections, and spoofed content before they reach your model. for 30 years email was a human interface. it's becoming an agent interface.
English
Superagent retweetledi
Introducing Infinite Monitor 🌐
- Monitor any situation
- Infinite amount of vibe coded widgets
- Each widget is it's on agent (you select the model)
- Runs locally and 100% Open Source
Github link and examples ⬇️
homanp@pelaseyed
I'm vibe coding an agentic OSINT/SIGINT app over the weekend. I call it Infinite Monitor 🌐 Each widget is its own isolated app with its own instance of claude code.
English
Superagent retweetledi
Before credit scores, banks had two options: lend to everyone and eat the defaults, or lend to no one and miss the market.
AI agents have the same problem with context.
There is no credit score for context.
That's what we're building brin to be.
homanp@pelaseyed
English
Giving your agent access to an email will give it super powers.
Brin.sh makes secures your agent from phishing attempts.
We just published a tutorial on how to connect brin with @agentmail
homanp@pelaseyed
How it works: AgentMail webhook → brin scans email → threats detected & labeled → your agent decides what to do Real-time protection against: • Credential harvesting • Agent manipulation • Data exfiltration Open Source, free to use: brin.sh/docs/integrati…
English
Superagent retweetledi
Most open source repos have a vouch list. "Is this person on the list?" Yes or no.
We just shipped a credit score for contributors. It's a GitHub Action that runs on every PR and answers a different question:
"What does this person's contribution behavior actually look like?"
English
I go through at least 40 plans a day, spend more time in "Plan" mode than probably any app at this point.
I feel like I need to do everything I can to increase what I call my Intent Bandwidth.
Wrote a piece about it:
x.com/pelaseyed/stat…
homanp@pelaseyed
English
@pelaseyed Hmmm, Intent Bandwith sounds like the name of the next big AI Co.
English
@pelaseyed Couldn't agree more. Go deep into uncharted territory. No one is safe, that's your advantage.
English
Superagent retweetledi
this is not a Grok-specific vulnerability. it is what happens when two safety layers have no shared context.
the prompt guard does not know what was generated. the classifier does not know what was requested. an agent with tool access does not know which instructions in its environment are hostile.
you cannot patch your way out of a design problem.
Jamieson O'Reilly@theonejvo
English
Superagent retweetledi
Coding agents are great, until they install something they shouldn't.
npm and other registries weren't built for this. Slow moderation, no agent-aware security, typosquats sitting undetected for months. When your agent runs `npm install`, it's trusting a system that was designed for humans who read changelogs and vet dependencies. Agents don't do that.
📦 Brin makes sure that doesn't happen.
It's the security layer between your agent and the registry, scanning for malicious packages, prompt injection, typosquats, and supply chain attacks before they hit your codebase.
homanp@pelaseyed
Your coding agent just ran `npm install`. You didn't review it. Neither did anyone else. Introducing 📦Brin the agent-native package gateway that blocks unsafe installs before they hit your codebase. Here's how it works:
English