AI Deployments Need Security Barriers

Mindlessly using Codex or Claude Code? You could also be mindlessly tossing your API keys to bad actors and giving away your personal data. Learn what you can do and how you can protect your personal data while still getting the most of powerful AI Agents. https://lnkd.in/dxRyKHbD

The AI pilot was supposed to eliminate the copy-and-paste between your systems. It quit after two weeks because the copy-and-paste wasn't the root problem. The real fix is the process that made the copy-and-paste necessary. Fix that, and the agent gets boring.

We spent a day trying to stop our own AI from agreeing to a price. We lost. Three times in a row. Here is the setup. People message us all day about houses we have for sale. Hundreds of messages. Most need a short, simple answer. So we built something to read them and reply. One rule mattered more than all the rest. It is never allowed to commit to a number. We wrote a filter to catch that. Then we spent the day attacking our own bot to see what got through. Plenty did. "We accept your offer." "Same as last time." Not a number in sight. Both are a commitment. So we tightened the filter. And it started blocking ordinary sentences. Things like "we can do Saturday at 11." Three rounds of that taught me something I should have seen on day one. You cannot make free-typed text safe by inspecting the words after they are written. So we changed the question. The bot no longer writes what it sends. It reads a message, picks which approved answer fits, and the sentence itself comes from our own records. Anything outside that short list goes to a person. It can still be wrong about which answer fits. It can no longer invent one. We stopped policing what the machine says, and changed what it is able to say. #PropTech #Automation #Operations #RealEstate #AI

AI agents do not become enterprise-ready through intelligence alone. They also need to survive failure, preserve state and resume work without starting over. In his guest post, our CTO Greg explains how AgentVisor maps Agent Protocol Threads to Temporal Workflows—making agent state, checkpoints and long-running sessions durable without requiring changes to the agent code. Thank you to Temporal Technologies for sharing Greg’s work. #GuestBlog

What if an AI agent could become durable without changing the agent code itself? In this guest post, Greg Haskins, co-founder and CTO at Manetu, Inc., explains how AgentVisor maps Agent Protocol Threads to Temporal Workflows so agent state, checkpoints, and long-running sessions can survive crashes and resume elsewhere. Read more: https://lnkd.in/de6nJfdp

"A number you can't trust isn't a result. It's a decision someone already made, wearing a decimal point." If you deployed AI this year and you're not fully sure it worked, this is the 30 minutes to watch.

"A number you can't trust isn't a result. It's a decision someone already made, wearing a decimal point." If you deployed AI this year and you're not fully sure it worked, this is the 30 minutes to watch.

What happens if one low-privilege AI agent gets compromised? Agents talk to each other and delegate tasks. One hijacked worker becomes a route into your most sensitive environments. In his latest blog, Jacob Abrams explains how to contain the blast radius. Check it out to learn how to protect your AI deployments: https://ow.ly/ivjT50ZOXOi

FBI alleged hack: 3 month unpatched vulnerability Why bother using AI when you can just use a 10 year old playbook together at the data you need

It's a two-way street. Since prompts and responses can be used to add to the data an AI uses, be careful not to load a bunch of false information and incorrect comments in your communication with AIs. Validate everything!