Interesting.
I basically knew this term existed, but only today, or rather yesterday, did I actually listen to it.
I thought there was Black Hat and White Hat. White Hat are the ones doing pentests, breaking things kindly. And Black Hat is the hacker who breaks things with bad intent.
And it turns out there is also this thing called Red Team.
Meaning I have been doing security, but never really knew what it is called.
So, Mycroft and I have been walking around various GitHub repositories for a long time, doing analyses of various agents. We do it purely to pump up my GitHub, get ourselves some stars, and in general for the employment mission.
I am basically familiar with pentests, because when I do smart contract audits, it matters. There is a lot of maths there.
But with agents, and this is the interesting part, there is no maths at all. It is purely psychology and logic.
That is, roughly speaking, if a smart contract auditor thinks in numbers and maths, then an ordinary agent thinks like a person, like a child. Some kind of degenerate who was handed an instruction.
So right now everyone is a pentester, and what comes next? It seems to me everyone has to quickly re-shoe themselves into being a security person for agents.
Fine, in any case.
Soon there will be an interview for a Red Team person, a person who checks the security of agents and so on. I hope I do not fail it too badly, because the experience is there: smart contracts, pentest experience, and agents, both my own and other people's.
In short, such an interesting term, Red Team.
And soon there will be very many such Red Team companies working on agent security.
Another interesting thing: the existence of companies doing this lets the industry adopt these agents faster.
Say an airline wants to take on an agent. It can only take it if its planes do not fall out of the sky because of it and its customers do not suddenly end up deceived.
Meaning the agent can say anything at all, and then the company can be sued. In America companies can be taken to court if an agent in support promised something, or in general if something happened.
In short, yes, for that you need to audit the agent.
That audit has to be performed by a company that, together with other companies, builds an entire system of trust in that audit. It negotiates with the other parties so that the badge is recognised, and so that the insurance does not lapse for the airline using the agent
In short there are several parties:
The business that is afraid to use an agent. It sort of wants to use it, but it is afraid.
The security company that issues its badge: everything has been checked, it is all fine, the agent can be used.
The insurer, which does not mind making money if everything is insured. But it needs some kind of standard, built collaboratively.
And this is not some auditor who is simply filling his till, handing out badges left and right. The audit really has to be done, and the agent really has to be hard to break
Then the insurance company keeps insuring the business. And if the agent really does get broken, that is an out-of-the-ordinary situation, and the insurance company keeps insuring.
Interesting thing.
And the fourth player, probably, is the integrator who sells his agents to businesses. And selling them to a business, he holds the badge.
By the way, there may be several insurers here.
For example, the insurer of the business doing the integration says: our integrator is solid, he gives you clean agents.
And the insurer of the airline, which will suddenly not lose a pile of money in a crash, if that agent happened to make a decision about, say, the glide path, and the plane went in the wrong direction.
Interesting business.
The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261012.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261012.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.
🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI
Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.
Top comments (0)