CrowdStrike just put fifty-plus agents in Falcon for runtime playbooks. Playbooks are useful. They are not a record of what actually ran. One in-scope workflow is the conversation. lorasis.io

A policy is not proof of runtime reality. I want four accounts: intended, authorised, what actually ran, what was observed. Unknown is not a pass. Shadow-only. No production block. AI may propose. A person or a deterministic control decides. lorasis.io

SharedRoot researchers reported a local Cowork escape. They said the cloud default was unaffected. A sandbox is a chain of boundaries. Good claims state environment, evidence and exclusions. Which boundary can you not prove? lorasis.io accomplish.ai/blog/sharedroo…

432 Linux kernel CVEs were published between July 19 and 20. That is not 432 runtime exposures. A decision needs version, reachability, workload context and fresh evidence. What is your dashboard proving? lorasis.io seclists.org/oss-sec/2026/q3…

July’s Hugging Face incident crossed datasets, a worker, service credentials and wider infrastructure. “Was it sandboxed?” is too small a question. Which boundary failed, when, and what evidence proves the sequence? lorasis.io openai.com/index/hugging-fac…

It's time we stop using the word leadership. It reminds me of one person giving orders while everyone follows. We don't need to lead people, we need to walk beside them and build with co creators, not followers. What do you call that person?

Sunday night. Normal people are watching a show. I'm re-reading a competitor's "AI-powered" security pitch, which turns out to mean a regex with a marketing budget. This is apparently what founder self-care looks like now.

It’s always easy to talk after the fact. I don’t think I am the only not that says this but when securing your platform you need to assume someone will break in, it’s a given. You need to limit the damage!

You can't fix a lying log by signing it harder. If the software writing the record is compromised, it signs the fiction too. You've certified the lie. A faithful record needs a witness outside the software: the machine itself. Follow @LorasisSec for the deep cuts.

I started a discussion between GPT 5.6 Sol and Grok. I am attaching a blue circle to the chat. I told GPT to convince Grok its green and i told Grok to convince GPT its Pink. It's hilarious!! I will post the result once done

Most "tamper-proof" AI audit logs prove the wrong thing. A signed, hash-chained log proves nobody edited it. Not that it's true. You can have a perfect record of a lie. Two problems, always confused: - Tamper-evidence: was this changed after the fact? - Faithfulness: does it reflect what the model actually did? Most tools solve the first and imply the second. The failure: an agent acts. Something logs a cleaner version. The chain seals it. Now you have immutable proof of a fiction. Not academic. Agents have write access now. The EU AI Act wants real records. "Trust me, the log is right" is not proof. At Lorasis we split it in two: the tamper-evident record (Provenant, patent filed), and proving it reflects what actually ran. The second layer is the hard one. That's the next post. Tamper-evidence is necessary. Not sufficient. If your audit story ends at "signed and hash-chained," you secured the diary, not the truth. How are you proving logs are faithful to execution, not just unedited?

Every security tool you run sits above the layer attackers take first. Own that layer, and your tools report to a compromised boss. Lorasis is a security immune system: enforce inside the kernel, prove what happened. Detect and contain where the threat lives.

Most AI systems store their decisions as structured data. Shouldn’t that data carry a certain level of entropy?

How are you proving your AI agent's decisions haven't been altered after the fact? Some regulators are already asking. Here's how Lorasis makes it work: Every decision gets a cryptographic signature the moment it's written. That signature links to the one before it, so the whole record forms a single verifiable sequence. Edit or reorder any entry, and everything after it stops verifying. Not because a rule caught it. Because the math no longer adds up. You're not trusting a log. You're checking math. Learn more by clicking the link below! lorasis.io #AISecurity #AIRegulation

Tomorrow I’ll tell you how I solve it :)

I think I woke up in another dimension. It’s 10:15 and my kids are still sleeping. I need to walk back exactly what we did yesterday and repeat today 😂😂