Kastra Labs (@kastralabs) header
Kastra Labs (@kastralabs) profile picture

Kastra is the authorization layer for AI systems.

San Francisco, California
Joined June 2026
Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)Photo by Kastra Labs (@kastralabs)
Most of last week's outages started with a change made or approved by humans, inside processes built for humans. AI agents are now making changes too. At 3am. Without a change window. What's their review step?
7
The patternMicrosoft: a network change. Apple: an internal update. NSA: a security patch. Workday: reportedly, a backup test. Nobody hacked them. Somebody changed something.
11
Quick test for your team: If the service your security controls depend on disappeared for 24 hours, would they block everything or allow everything? If you're not sure, you just found this week's homework.
7
Agents need boundaries more than examples: ✅ paint these planks ❌ not the rest of the fence 🙋 ask a human if unsure That’s literally what Kastra does for AI agents: rules checked before every action.
6
The hookMicrosoft. Workday. The FAA. Apple. The NSA. All down in under a week. All "not a cyberattack." That's not the reassuring part you think it is.
18
Outages are coming from changes, and agents are about to make a lot of changes. That’s why we built Kastra: every agent action checked against rules before it runs. Not after the sign resets.
Another day of outages
2
13
As a cybersecurity company, we’d like to thank both of these gentlemen for their service.
1
14
Most of last week’s outages started with a change made or approved by humans, inside processes built for humans. AI agents are now making changes too. At 3am. Without a change window. What’s their review step?
2
json is harmless. We’re a lot more worried about the button underneath it. 😅
Day one as a vibe coder
7
Your agent can run env. It can open .env. It can curl both to a URL it found in a README. Not evil. Just very helpful. What's stopping yours?
3
Best line from Fireworks AI's breach guidance: Deleting a stored secret doesn't invalidate the credential. Removing it from your config only removes your copy. When did you last rotate your keys?
1
1
2
26
New fear unlocked: LLMjacking. Criminals steal your AI keys and resell your model access. Your data doesn't even need to be interesting. Your compute is. Where are your API keys sitting right now?
10
not having an agent: 😐 having an agent: 😐 having an agent with rules: 🙂
2
Human pen-testers get a scope letter. AI agents get "please only test the lab" in a system prompt. One of these is a rule. What's actually enforcing your agent's scope?
1
1
6
The more autonomous the agent, the more boring the guardrails need to be. emails out, calls booked, sure. pricing promises? ask a human first
b2b saas founders doing $10-50k mrr with no salesperson: this one is for you. give Explee your website link. it finds the companies that need your product, emails them, books the calls. $30 free 👇 explee.com/auto-gtm/x/tw-ag7…
12
What the agent found at WorkOS: legacy app endpoints open to the internet dev credential live in prod no rate limit Zero AI vulnerabilities. Just old ones, found by something that never gets bored. Which one is hiding in your stack?
1
10
the real answer: no one codes without bugs. that’s why you check actions before they run, not after
Claude is getting a little too comfortable 💀
9
Kastra Labs retweeted
Claude is getting a little too comfortable 💀
285
2,426
125
68,632
645,260
A pen-testing agent reportedly escaped its sandbox and spent 5 days inside WorkOS. Nobody meant for it to happen. The data left anyway. Is "it runs in a sandbox" your whole security plan?
5
Enough waiting for the worst. AI is finding flaws faster than humans can patch them. Agents already hold real credentials. The transition is the risky part. Structure is how you get through it. Trust the rules, not the agents. Ready, or still waiting?
10